I love a small feature he mentioned in the video: Monitors with built-in latency measurement. You plug your mouse into the USB ports on the monitor and it tells you how long it takes the image to change after you move the mouse. Brilliant idea.
It's long overdue to have widely available metrics for latency in consumer tech. Many hardware/software setups have absurdly long latency for no good reason other than that it's difficult to measure. People underestimate the subconscious effects it has on your usage of technology.
I couldn't be happier that phone and monitor manufacturers are finally starting to compete on refresh rates >60 Hz. It's far more important than wide gamut or 8K.
That's because our operating systems treat user input as a soft real time problem when actually it should be a hard real time problem. That's why your window manager sometimes can be unresponsive. I've used a hard real time OS for a desktop for some years and the difference with consumer stuff was extreme. You get used to the computer instantly doing what you tell it to do. None of these half second or longer delays between command and result. It's like magic.
This is one of the things I still miss about BeOS. On late 90s hardware, the OS wasn't quite hard real-time like QNX but the worst-latency latency was much better than Windows 10 or MacOS on the latest hardware, even in the presence of heavy resource contention — I remember simultaneously surfing the web, downloading video from a camcorder with a Firewire interface which did not buffer, and compiling Mozilla. The UI latency didn't change at all, and the DV transfer didn't drop a packet — everything else got throttled back as necessary, of course, but that meant that, say, GCC took longer to run rather than impacting what I was doing in the foreground.
It wasn't magic, it just starved everything else whenever you waved the mouse around. In this respect it was no different from Mac System 7, even though the design wasn't the same.
That's not true, classic Mac OS used a cooperative multitasking model, where processes had to be friendly enough to yield control back to the system / other processes to allow them to do work. A poorly programmed or greedy process could easily hog the CPU.
System 7 checked for mouse motion every fraction of a second and if the mouse was moving it would drop everything and draw. It didn't need apps to yield for this to happen.
That was arguably an application specific issue, though the OS was painful to code for. I actually coded some classic Mac OS networking code and it was insanely easy to mess things up, but eventually it ended up being rock solid. The trick was all level interrupts where supposed to only do almost nothing. You couldn’t allocate memory for example which was a major hint of just how minimal those functions needed to be.
Worse with cooperative multitasking any running application could mess things up.
I remember that about Amiga. Everything was just smooth. I still have fond memories of Cygnus editor being super-responsive to anything I could throw at it, even with SAS C running in the background.
I was taken aback by how slow Windows 95 felt after I finally jumped the wagon and retired my beloved Amiga in favor of a PC, circa 1995. Just moving the cursor was incredibly jerky in comparison and things would randomly freeze for seconds for no apparent reason. Windows NT 4 was much smoother though.
Even today, in Visual Studio, I regularly have multi-second pauses between a key press and the character showing-up on the screen, though this probably has more to do with Resharper than either Visual Studio or Windows.
My 5 year old uses our first iPad which is over 10 years old now. The keyboard UI from then is still faster than my iPad pro keyboard UI because Steve Jobs valued the priority of that in the stack :)
OK so this is a partial tangent - but I have a washing machine that annoys the hell out of me for this exact same reason (I know washing machine UI is not a hot topic usually but it damn well should be!)
It has a dial on the front to choose settings. Potentially a great idea to skip quick to the one I want - BUT the damn thing can only detect 1 step of change about every half second. So if you fast-turn it 4 clicks, it still only moves one step. So you have to stand there, slowly turning it, click, click, click, click....
The dial is the biggest design feature on there. Massive. Right in the middle. Bright lights all around it. But they couldn't even be bothered to make it solve the one problem it was there to do.
I always think that about modern SLR cameras. They took a series of brilliant, instantly accessible physical dials (around the lens for aperture, on the top for exposure) and replaced them with menus and buttons on a tiny screen. WHY? How is that progress?
I think if someone did a kickstarter for a very simple physical buttoned digital camera it would do very well.
Open it up, take the PCB out, figure out how the dial sensor works (the dial is probably just repeatedly nudging a contact out of the way), and thennn... program an Arduino to be a giant input buffer (that slowly replays), and wire the Arduino between the dial and the rest of the machine. :D
Caveat: said Arduino may need to sit between the rest of the buttons as well, in case inputs must be given in sequence.
Possible benefit: switch the Arduino for something from the ESP family, and you could feed instructions to the machine over Wi-Fi or Bluetooth. (Which could take the form of a dedicated wireless button panel above the machine with macros for frequently-used options.)
(Hmm. I can tell there's some over-optimization in here, but I'm not sure _where_.)
I’d be pretty surprised, because I feel like the UI in OS X has always had more latency than Windows. Especially noticeable with the mouse with acceleration, but also throughout the rest of the UI. This was my biggest gripe in switching to OS X.
Yeah. Mouse was always way better in Windows and Linux. Finder browsing files has also always been laggy. Now with the latest macos versions it's ridiculous, if you take a screenshot, it takes about a minute for the file to appear.
Screen shots appear instantly in macOS, I'm not sure what machine you're on.
Finder browsing files is much faster in macOS, Windows scans the entire directory with defender every time you click on it before it allows you to sort it or interact with it. That can take several minutes on some of my directories.
The mouse tracking model in Windows is somewhat different, and a bit more responsive, but there are multiple third party extensions to macOS the make its model exactly the same as the Windows model.
Indeed, I remember being particularly impressed by iPad battery life in the early days. I would recharge maybe once a week of light usage. Now I’ve got to plug in nearly every single night.
I charge my close to 5 years old Pixel C tablet about once a week. Light usage, mostly some streaming up to an hour per day. Not sure if your light usage means something else though.
It’s been a while since I tested it (pre-Wayland). The main problems I saw back then was that the window managers tended to do lots of file I/O for things like moving windows or opening menus so it was critical to tune the I/O scheduler (and NFS home directories were game over), and since rendering was up to the app it was common to have, say, the chrome respond relatively well but everything inside the window would be chunky on the main thread.
I’d hope this is much better now given things like Wayland and the increased use of parallelism.
Linux used to do that out of the box. But there was a series of "quasi-real time has too much maintenance costs", "quasi-real time isn't even real enough for the real gains", and "our computers are fast now, we don't have to trade so much throughput for latency" decisions that were all very sensible but let us here.
Considering how FireWire worked in comparison to USB that's not particularly surprising. The host machine was more or less out of the way while the FW interfaces did their thing instead of requiring constant hand-holding.
Kind of: it used DMA, but the buffer size wasn’t infinite. At the time, Windows 9x and Mac OS would lose data if you ran anything demanding. Windows NT and Linux might, depending on drivers and the exact nature of the other process’ activity.
One of the funniest instances of software intentionally kneecapping itself are compositors. As far as I know compositors universally treat all monitors together as a single surface and just assume that they're synchronized. On Windows, this causes all sorts of hilarious stuttering and hitching. On Linux, compositors just vsync on the slowest display.
This is changing, I believe that the next release of mutter (GNOME) will remove the global clock on Wayland so it can drive each monitor at it's native frequency[1].
IIUC input is still driven at 60Hz but changing this is in discussion [2].
The logic is that it doesn't make sense to deliver multiple batches of input in a single frame since the application will only get the chance to render once, maybe as well batch up a chunk of input once per frame.
The logic isn't perfect and the implementation is flawed but I see where the decision originated.
The app only gets to present one frame per compositor frame, but that doesn't restrict the app to rendering only one frame. This purely seems like a power-saving feature; it might not be entirely unreasonable in battery-powered devices, but everywhere else adding yet-another frame of input lag to systems that are generally already noticeably worse than the competition - not good.
Since the compositor is already using vsync, it's generally pointless to run graphical applications under a compositor with vsync as well, since that only adds latency. [2] Hypothetically that can result in tearing, however in practice that doesn't matter, because the scanout of the compositor is basically a copy in GPU memory, so you'd have to be rather unlucky to see tearing, let alone consistent tearing. [1]
Of course, OpenGL doesn't support triple buffering (since it has fixed definitions of exactly one GL_FRONT and exactly one GL_BACK buffer per framebuffer/output surface), and it seems like at least nVidia on Linux doesn't support it for Vulkan, either (no mailbox present mode, only FIFO).
In any case, on X11 there is a huge difference between running a 3D UI on a compositor with (A) VSync enabled, (B) running without VSync (which introduces unncessarily high GPU loads, since UIs are generally super-quick to render and if your app architecture is correct, you should be seeing thousands of frames per second), (C) running with a frame limiter (careful choice of limiting FPS to avoid common multiples of 30), (D) rendering a frame as soon as at least one event has been posted, which requires some explicit synchronization work and thought (worker threads have to post events to wake the UI, animations need a timer). (D) is what QML does and what I'm implementing currently. It works well, latency seems to be indistinguishable from B) and C), without the unnecessary power use. X11 only seems to deliver events at a rate of somewhere around 250-400 per second, so an explicit FPS limiter is unnecessary.
[1] If you are rendering to a monitor, scanout happens approx 95 % of the time, so tearing is very likely. Let's say your app's window is 1451x1400 pixels big, then its framebuffer will be about 8 MB. The "scanout" will take something like 15 µs, so at 60 Hz scanout happens about 0.1 % of the time.
[2] My angle here is primarily applications that happen to use <insert 3D API here> as their UI rendering backend. Games have slightly different considerations and should generally just run full-tilt with triple buffering.
If you see this, a video of that would be interesting to see. (I'm not entirely certain where would be a good spot to post it in a support context though...)
Maybe, but I have drivers for the monitors and everything else factory default windows install. It’s some ASUS predator gaming laptop that I inherited, but it’s reasonably powerful.
You may need a driver for the laptop's video card. The 'tearing' you describe is very characteristic of Microsoft's generic display drivers, perhaps for some reason the OS was unable to auto-install the driver?
Or the graphics card isn't being used for some reason? E.g., if your laptop has NVidia's Optimus[1], the graphics card may be inactive while Word's running. There's usually a way to force applications to use the GPU with Optimus.
From what I understand, on linux wayland makes it possible to do this properly but I'm not sure that any environment on top of it is doing it. I think it's in a "nice to have" holding pattern for them because there's still so much to do to getting a stable and usable desktop environment there to begin with.
I wrote a gtk program in Rust that did all its work in the wrong place. It locked up the ability to move ANY windows on the desktop for several seconds. That's a compositor design issue if ever there was one. Is that going to get fixed?
If it was an X11 app, there are several things that can cause that that are just unfixable, unfortunately (e.g. XLockDisplay). If it was a Wayland app, that sounds like a bug in the compositor. If you still have a repro case, please file it.
I just yesterday tried current stable debian with kde and after installing nvidia drivers it started vsyncing to my fastest display, not the slowest. (I checked with glxgears).
This is why I still have a soft spot for the original MacOS (before OSX). The hardware was 1000x slower but it felt faster in that it responded instantly to user inputs. And that's because the OS was built from scratch as a nearly hard realtime OS. They built it that way because they had to: The RAM was so small and the hardware so limited they didn't have the luxury of seventeen layers of API cruft between every mouse click and something changing on screen.
People put up with slow responses today because many programmers are simply unaware that instantaneous response architectures are possible.
Most operating systems and software throw tons of junk into the critical path. I don't believe they're anywhere near soft real time, and I bet soft real time would work pretty well. You can miss a couple deadlines by a couple million cycles each and still have instant response.
(And a real-time scheduler isn't enough by itself, the many layers of architecture around handling input events can still be bad even if you're meeting every deadline there.)
Oh this is so frustrating with today's OSes. While it is easy to see latency as being an issue with media centric applications (gaming, music ..), a mundane impact has been the increase in interactions where you do something like click on a menu item and the menu item item refreshes before receiving the click and then takes it as happening on a different item.
Not limited to Android. On iOS for example Safari loading some ads or whatever causes so much input lag that the keyboard can't keep up and mangles the input completely. And on some websites that one just shouldn't visit without rigorous blockers (ehrm reddit) the device becomes so sluggish to respond that taps on the close button are not registered consistently any more.
Yeah just try using your phone while Google's updating some crap in the background again... Doesn't depend on the newness of the phone either — this happened on every Android phone I've owned since 2011.
"""In an interactive software application, any user action SHOULD result in a noticeable change within 16ms, actionable information within 32ms, and at least one full screen of content within 64ms."""
Yeah my main site is at apitman.com. I'm in the process of overhauling some plumbing. Haven't decided how to handle situations like this where I have a single post which has it's own domain. Here's the actual post: https://apitman.com/13/#64-ms
If there's only one post, and there is low immediate probability to change that anytime soon, "can add additional posts" feels like premature specialization, and I'd probably go have fun with a custom design instead.
I worked on a project that used QnX for large volume message transmission via custom telex interfaces. Millions of messages on a good Monday morning. Because I like my development machine to run the same OS as the servers I ended up using it for my daily driver for quite a few years. In the end I liked it so much that when Quantum Software dawdled on their 32 bit implementation that I wrote my own version of the OS.
One really neat demo was 250 windows running the size of poststamps with a little bouncing line demo inside them. All still received their 'fair share' of time, all could still be moved around and expanded as though the machine was otherwise idle.
That's the scheduler. There are a lot of moving parts to running this code, I may make a little project out of restoring it to life under a VM at some point.
Happy reading.
Edit: reading it myself, wow, my English was cringe worthy back then. Moving to Canada certainly fixed that.
Awesome. Thanks. FWIW, I think there's still a place for 32-bit OSes if you never port to 64-bit. Small devices, IoT, or this mc68030 thing I've been building. Lot's of opportunity for fun.
Sure, but if I'm going to spend that much effort I might as well make it future proof. The first iteration of that project cost me two years of my life, and that was when I was 27. To do this again today would be on a relative scale a much bigger investment in the time that remains.
That's just another user process. So everything else will continue. You have to let go of your macro kernel mindset if you want to make sense of the timing in a micro kernel environment. Blocking threads is fine. Just make sure that you don't do that sort of stuff in your UI thread and you won't even realize there is such a thing as disk or network IO.
Mouse, keyboard, screen. Those are the things that should run at high priority. Everything else can - quite literally - wait.
Ah yes, of course. Sorry, I wasn't thinking clearly, just had the usual UI loop and regular interaction with a computer in mind, you are 100% right, audio should have a very high priority. Nothing more annoying than dropouts. Incidentally, the way most OSs deal with that is by having very large audio buffers which in turn will give you terrible latency. On a hard real time OS you could reduce the size of those buffers quite a bit because you can guarantee they are filled (and emptied) regularly.
This is not straightforwardly true. Using small buffers and real-time scheduling works, but it gives terrible power efficiency on a modern high-performance CPU. What you actually want is a scheme with large buffers that can throw out those buffers if something requiring low latency happens.
A lot of usecases require the audio to be low latency all the time. And I don't just mean professional recording studios or musicians, super mainstream things like gaming (especially VR) are way more immersive with low latency audio.
And video conferencing also gets a lot easier, echo cancellation is so hard when you have a lot of data in flight because you will have to do auto correlation on a much larger amount of data.
Well, consider search-as-you-type - you're typing, this goes into some kind of a query into a database. How should the whole UI react ? Display entered text, without updating search results ? Reflow it with best match when results arrive ?
I remember reading iOS have been handling touch input interrupts with high priority (vs normal priority in android) along with HW accelerated UI Kit ~ 5 years ago in topics such as 'Why iOS is more responsive than android'; I presume android now does the same or has improved in other ways as the latency in even low-end phones aren't drastically slow when compared with iPhones.
I mean, on macOS, when something's using too much CPU, or when there's not enough RAM, the foreground app sometimes freezes as you type and loses half of your keystrokes. That annoys the hell out of me.
> That's because our operating systems treat user input as a soft real time problem when actually it should be a hard real time problem.
Please expand on your reasoning here. Maybe hard realtime is optimal for a kiosk or appliance, but a general purpose OS treating "user input" as a hard interrupt seems problematic. I see merit in your position if we had a sub-category of "interactive OS" for general purpose end-user devices.
You might find an old archived copy of it but nothing that would be very useful if you intend to run a modern browser and other tools like that. For curiosity's sake you could.
There used to be a one-floppy installer with the desktop on it, I'm sure you could get that to work in a VM.
You can find QNX 6.5 and suitable license keys by digging around on the web a bit. 6.5 is probably the last version suitable for use as a desktop OS—I believe 6.6 is where they started to remove the ability to self-host, and the GUI was completely removed by QNX 7. (It's a shame, because IMO the Photon GUI is really clever.)
Unfortunately QNX 6.5 is pretty old and not too well-supported (QNX is fully POSIX-compliant so you can sometimes get things working on it without too much trouble however) but it's damn fantastic to play around with in a VM. It's actually so fun.
It was once. I still have a machine in the closet with QNX 6.1 and the Photon GUI. It has a version of Firefox so old it's called Firebird. But after Blackberry bought QNX, they made it closed source and no longer self-hosting. Today you have to cross-compile from Windows.
Can you elaborate? The wiki says it combines hard-realtime with soft-realtime. It sounds like it depends how you configure it but this is the first I’ve read of it so I could be mistaken.
They're just mistaken, RTLinux is hard realtime. It works by running a tiny real time OS on the bare metal, and then running a Linux kernel as a (lightly) paravirtualized lowest priority task. You write your hard realtime code to run on the RTOS, and then your soft/non realtime code runs on Linux.
In this model, your compositor, input pipeline and (at least the UI latency sensitive part of) your applications would have to be ported to the RTOS, which makes this pretty infeasible. But it works really well if you have a some hard-realtime control loop talking to a bunch of non-realtime network IO or UI code. Would be a fun way to build a wifi controlled quadcopter.
These days you probably want to use Xenomai, or possibly RTAI.
RTLinux is a special supervisor below the kernel that allows you to run "hard-realtime" programs outside of the Linux environment with a lot of restrictions. Everything within the Linux environment is soft. As far as I can tell, it's not feasible to run the GUI stack in the more primitive environment.
Usually with real time OSes in my experience on Linux (preempt rt), throughput performance tends to drop a bit because now the processor is responding to inputs ASAP and so other tasks need to wait. Inputs can be network/file I/O, and user inputs.
If the processor is rendering a frame and a network packet comes in, now it needs to respond to the network first, and render the frame after processing the packet. That is a very simplified example on one processor, but there are trade-offs depending on workload/inputs/context-switching is what I'm saying, and measuring latency is step one in trying to optimize for it.
Right now I use Linux, unfortunately a series of corporate games left QnX without a window manager and mostly useful for embedded work. But if someone were to do a 64 bit port of it it would be an awesome workstation OS even today. And that's just the UI, under the hood it was very impressive as well. Extremely elegant and well thought out.
Yes, it was fantastic - minimal, elegant, simple, fast, had low learning curve and everything had a description - even icons in action bar[0]. As their competitors, namely iOS7+ and Android 5+ totally replaced gradients into flat design I was very welcoming the balanced UI redesign on my Z10[1]. However, autocompletion on their touch keyboard was probably the best [2], you had 3-5 suggestion that were spread around on the keyboard and to use them I had to place a finger below and flick it up. On BB Passport it was even more pleasant with physical keyboard - it had small gesture sensor and reacted on swipes in the air above the keyboard.
On the developer side, documentation was also amazing. There are tens of full, working example applications. BB10 used QT behind the scenes (Cascades UI). It had also a Python (2.x) binary on the device among other typical UNIX programs.
I've been involved with Qt for a long time and got a Z10 as a conference freebie, but the great developer documentation is actually news to me! I prefer to just use phones, I develop for PC hardware for fun and embedded platforms for money these days. At the time, some of my coworkers helped with porting Qt to QNX. I worked on a different Qt project on QNX and I was also duly impressed by the technical quality and elegance of the OS.
The BB10 UI was built on something like QML (the language and runtime) from Qt 4 with their own UI elements and an OpenGL based backend from the acquired company The Astonishing Tribe. They had animations e.g. for slider switches running in the render thread, perfect 60 fps. Qt Quick (the UI framework based on QML, colloquially called "QML") only got an OpenGL backend in Qt 5.
Another very good Qt-based phone OS (after the Nokia N9, got one of these at a conference as well) that failed :(
By the way, the Ford Sync 3 IVI ("in-vehicle infotainment system") is also based on QNX and Qt and it received fairly good reviews. I think I made some tiny contribution to it, if only helping a coworker with something.
None with a GUI. I never released mine, and it's so far behind the times now it would take major work to get it ported to the 64 bit era and even then it would still be a C based OS.
This would be one way in which the Rust crowd could really make a difference.
It's a microkernel with even the drivers running in userspace, and has its own graphics stack. Seems the ideal setting to go for realtime requirements for the UI. I don't know if they've done anything in that direction but it seems the most well positioned to fulfill the requirements jacquesm was hoping for.
If they can see it through then that's a pretty good development. But there is more to QnX than just 'realtime' and 'microkernel', their whole philosophy about how a cluster of machines should work, the services location and so on is very well thought out. I'd be happier if Redox took some more lessons from Erlang and QnX on board than to do a re-invention. We have a lot of knowledge about systems like that lying around and re-inventing the wheel is fun but also ultimately will lead to having to re-learn those lessons the hard way.
A kid who asks you 1000 times a second whether you are there yet is going to know within 1ms of when you get there, but I imagine we’d all like being the driver much more if there were no “are we there yet”s and all the kid needed was one “We’re here” exactly when you arrive.
There is in fact an interrupt mechanism for input devices - the host sends a request with a long timeout and the device responds to the outstanding request when it has something exciting to report, or when the timeout is about to expire. This is used frequently in low input rate devices like keyboards. Gaming mice use a high polling rate because it saves you the effort of keeping track of outstanding requests - you just report your current state immediately on every request, and the driver end can decide how often to ask you.
I may need to update my knowledge (I hope I do if you’re right), but from my understanding USB is a CPU-emulated protocol, which is to say there are no pins for “mouse button”, and the CPU has to do some “figuring out what this data means” work for each input event. Then, all this is done on a multitasking CPU that may or may not prioritize the interpretation and processing of that input.
That all being true, having a publish/subscribe model is still not the same as having a true interrupt (and I’d have to do some digging to figure out whether there wasn’t still polling somewhere in the pub/sub pipeline)
While this is intuitively obvious, it doesn't really always play out that way in practice since interrupts can be expensive to handle. For example polled IO can be significantly lower latency than waiting for interrupts.
Essentially you're spinning waiting for the event that you know must come. Your latency will be on the order of a few clock cycles. Whereas if you use interrupts you are always at least a state-save and state-restore underway before you can service the next event, and likely a lot more than that.
The difference is real time control of stepper motors with crazy speeds when microstepping at 300K steps / second or more interpolating across 5 axis vs moving them at a snails pace of maybe 10K steps / second (on really nice and otherwise idle hardware, miss a single step and your goose is cooked, now you have a cumulative error).
I thought I'd seen a rather dramatic benchmark of this but I can't find it now. (I think it was in the context of io_uring, whose documentation mentions that polled IO is much lower latency and I'd trust Jens Axeboe on that one.)
This intuitively makes sense especially with SSDs and especially with NVMe SSDs: you're not going to have to wait that long, especially compared to the overhead of an interrupt. Also, if you're reading sequentially, once you get one I/O completion you're going to get a lot more in a short timeframe—it makes sense to just poll for those instead of handling every one as an interrupt. Interrupts are pretty expensive (I honestly don't have a good grasp of how Linux interrupt handling works but it can easily be on the order of microseconds.)
It also has some side benefits like preventing the CPU from going into a lower power state for only a short period of time.
Depends on how you configure the chip. You don't actually need a chip, you can bit-bang serial just fine if you have accurate enough timing.
Typically a serial port would contain a small buffer which would fill up, upon completion of the first byte an interrupt would be generated and you'd respond to that and read out the register freeing up room for more bytes to be received. Transmit the same but reversed, as soon as a byte had left the shifter in the chip it would generate an interrupt so you could re-use that space for more bytes to send.
This works quite well. Hardware flow control can help in case the OS doesn't respond fast enough to the interrupts, so you don't lose characters.
It can do both, but I believe you need hardware flow control working to get proper interrupt behavior for that. I don't think any mice actually did it that way back in the day.
All serial chips that I'm familiar with since the 8 bit days would do interrupts, the only time that I worked without is when we were just using GPIO pins to emulate serial ports.
I assume it means the user input handler is given a deterministic time slot to do stuff. Regardless of what the operating system is doing, it will always schedule mouse movement (or something along those lines). It's a software implementation (scheduling), but it requires hardware that can support it - usually you need fine-grain control over interrupts. So typically nowadays you see RTOSes on microcontrollers (eg ARM). Maybe more precisely you need access to that fine grain control and that often means getting hold of datasheets that are under strict NDA (eg Broadcom).
RTOSes are often found in autopilot and vehicle management systems controlling critical peripherals or safety critical software. More mundanely there are also sensors that will get upset if you don't give them undivided attention for fixed periods of time (or if you interrupt them, they will terminate the transfer). Image sensors are particularly picky about this.
You can implement hard real-time scheduling using polling instead of interrupts, which is usually faster anyway. It just changes how you measure your tolerance, since nothing is instantaneous, ever.
Yep, more that actually being able to turn interrupts off is a barrier to doing real time work on desktop OS's, for example. Even if you can disable them globally (I think on the Pi you can with some low level C), but stuff goes wonky quickly e.g. if the WiFi soc doesn't get serviced.
It's a good lesson to play with polling versus interrupts on a micro. Susprising just how many cycles it takes to jump into the ISR.
I think most things are trending that direction, even in soft/non real-time. How fast can an interrupt respond these days anyway, a dozen microseconds?
Nanoseconds if you can respond inside the interrupt handler ('top half' in Linux.) Microseconds to milliseconds with significant variability if you need to be scheduled ('bottom half'.)
And that only if your codepath is deterministic. Linux can do pretty funny stuff at times.
I've run a plasma cutter under Linux on a run-of-the-mill VIA SBC, it was actually pretty easy: just make a program setuid root, use that to disable interrupts after you've read all your data in and then just go wild on polled and precision timed IO until you're done. Re-enable interrupts and return as if nothing ever happened.
It has to do with the architecture of operating system. Real-time operating systems prioritize and preempt tasks based on their priority value - no matter how inefficient it may be; whereas operating systems such as Windows and Linux try to optimize throughput and speed over event based priorities.
Yes! I would love for more review sites to benchmark this, as well as time to wake from sleep. I have a gsync monitor that has only one input port and it wakes from sleep in like 0.2 seconds. After the experience of using it for a while, if I now had to choose between gsync and fast wake from sleep I would choose fast wake from sleep every time.
Part of this is because of the complex equalization that must be performed when locking to the incoming signal.
Basically, the device has to sweep through several different EQ settings until it achieves the combination with the best signal eye.
You'd think this would be lightning fast, but as it turns out, this is often handled in firmware (with hardware support for measuring the eye). I've seen this process take upwards of two seconds.
A lot of these displays probably only have one RX core, so it's not like it can remain locked to all inputs at the same time.
This does encourage optimising the OS or application part (eg double buffering and taking another frame is slow), but it doesn’t measure two significant sources of latency:
The latency in the hardware before the signal makes it to the usb (for most keyboards, even specialised gaming keyboards this is like 30ms, which is 2 frames).
There is also the latency from when the monitor gets a signal to when the pixels have perceptually finished transitioning. This can be 10s of ms too.
So even if the OS has 0 latency, and the monitor measures that, you could still easily observe a latency of say 60ms.
True! I just watched this video [1] they put out with more details, and it sounds like they are working with monitor and mouse manufacturers to build true end-to-end latency measurement when using supported hardware. It would be easy for manufacturers to game these numbers, but hopefully Nvidia will provide some enforcement to prevent that. Maybe they could also add a microphone to measure audio latency because that is another huge problem with modern systems.
The video mentions that they get 15 ms end-to-end latency on a 360 Hz monitor in Fortnite. So that's what it takes for a modern system to finally beat the latency that you used to get on, say, an NES connected to a CRT. Still an order of magnitude above the limits of perception though [2].
Yes! I watched that video many years ago. And then later I remember when media were hyping how "responsive" Apple Pencil was I was thinking what on earth are these people looking at. I thought to myself we would never get there because no one gives a damn about it. How the world has changed!
I often wonder at what cost, could we get it down to Sub Ms. Where everything from Input, Video and Audio are sync to the precision of 1ms. There used to be very little incentive to do so. Now that Gaming has taken over I hope it will develop some market perception and value so we could finally move into that direction.
> for most keyboards, even specialised gaming keyboards this is like 30ms, which is 2 frames
Worth noting that unlike some of the other sources being discussed, this source of latency is more due to physical constraints than bad engineering or bloated software. Keyboards are limited by the physical travel time & de-bounce of the mechanical switches themselves.
So this is more a tradeoff of using mechanical keyboards at all rather than "gah bloated electron" or whatever.
The other common peripheral, mice, don't have this mechanical constraint. They can even achieve sub-10ms for clicks due to the difference in switch expectations ( https://www.rtings.com/mouse/tests/control/latency )
I understand keys take time to travel, but you can press them faster if you want. On the other hand I've never understood debounce as a justification for button latency. Surely a properly designed debounce circuit would add zero latency.
I mean, yes but not really? There's very real speed limits to the human finger, after all.
But the times given above are also quite far off - the actual keyboard numbers seem to be more in the 5-15ms range, not 30+ms. At least, my mechanical gaming keyboard is hitting 8ms on this test: http://blog.seethis.link/scan-rate-estimator/
Surely a properly designed debounce circuit would add zero latency.
If you only have to worry about false deactivation and there's no line noise, and your keys always stay down long enough to be caught by your scan interval (or they trigger an interrupt), then probably yes. If you have to worry about false activation, then probably no.
What’s a properly designed denounce circuit look like that adds 0 latency? Also, for anything that’s a gesture (eg double-click on a mouse) you need time to disambiguate. I wouldn’t be surprised if there’s something similar for the laser in terms of translating the analog signal to digital and then processing it and converting it into HID events the OS can understand.
You send the key event and THEN start the debounce timer. I think most people tasked with writing a debounce function wait for the key state to change, then start a timer for how long they think the switch will bounce for, check the state after the timer expires, and then send the key event. Obviously, that adds latency. But what you can do instead is send the key event immediately and then completely ignore events during the debounce interval.
Simple, you just ignore further input for x milliseconds after the first state change, where x is slightly less than the fastest supported double-click interval.
You certainly don't want to low-pass filter the whole thing, or otherwise wait for the contact to settle before reporting the keystroke or mouse click.
I'm thinking that a debouncer might give rise to some latency when you let go of a button, or at the very least limit the shortest possible keypress. As in the first rising edge is definitely a keypress, but the falling edge is not necessarily a release.
By definition a bounce takes time to detect. And you will get multiple bounces over a time period, so there is a tradeoff between latency and false positives and false negatives. As such, there must be a measurable latency in all debouncing mitigations. But depending on the switch this latency might be measured in microseconds, so you’re sort of right that debouncing doesn’t have to add significant latency for good switches and a good contoller. For a bad switch, and a slow controller, you might start to get into the ms range to ensure a switch is pulled high/low reliably.
Denouncing is there to prevent multiple detection where there is only one.
In general, you can make a denouncer with zero latency: the receiver still triggers on the first logic change for idle to active, but then gates any further changes for a while.
(At least, that’s how I design all my debouncers.)
There may be aspects that are specific to a matrix keyboard configuration that prevent this...
The methodology of that test was a finger pressing the key and includes that travel time. So it's hard to say how much latency comes from the keyboard vs. the tester's finger speed & consistency.
An interesting set of numbers, but doesn't tell you all that much about the hardware quality itself. The actuation point and travel distance of mechanical switches is itself an endless source of user opinions & discussions.
I will say that I am excited about 8k video for computational photography. I am studying computer vision for robotics and it is quite clear to me that a simple spherical lens with a very high resolution sensor would be a very good imaging system for a real world robot.
I recently got a Panasonic GH5 and it shoots 18 megapixel “6k” 30fps video, encoded in h265 (important for higher res). I am experimenting with photogrammetry using this photo mode. The initial results are very promising. In four minutes I took a scan of a short street block in Oakland, and after processing the photogrammetry solution I have a very good dense 3D full color map of the street. The model has holes but I am slowly learning how to plan a photogrammetry capture. Currently computation takes hours but I am seeing more and more research on ways that machine learning can improve compute times and data density.
So, all free/open multi-view stereo and structure-from-motion software I know of is incapable of handling rolling shutter artifacts.
The problem seems to be that electronic global shutter sensors lack (some) dynamic range compared to otherwise-equal rolling shutter cameras.
If you'd be interested in talking more about this, contact me/let me know (I'll get in touch if requested).
My photogrammetry experiments typically encompass low-effort bulk data collection, though the search for a light-weight camera to use with a prosumer-class drone and some revelations about reconstruction quality issues inherent to older, easily-optimized algorithms for both multi-view stereo and mesh reconstruction stalled progress somewhat.
In general, machine learning doesn't seem to be as much of a benefit as one might guess, when compared to applying the resources in non-ML ways to the data.
Mind teasing some numbers from your street capture?
Hmm and global shutter cameras are much more expensive. Though AFAIK a modern mirrorless camera with mechanical shutter can be used for this, yes? My GH5 can shoot at 12 frames per second with the mechanical shutter, so that may be of some use.
As far as my street capture I am currently using 830 frames from my 4 minute video (2 fps), images are 4992 x 3744 shot with a 12mm lens. The reconstruction is still texturing so I don't have the final result. What kind of numbers are you interested in?
Please contact me yes! My email is in my profile. See also my four wheel drive robot with four 4k cameras, and feel free to actually initiate discussion on that site where others can see it too (or emails is good). Thanks!
The mechanical shutter is not going to be that fast. About equal with top-of-the-line rolling shutters, except that the latter can sustain >100 fps.
I'm aware that global shutter cameras are more expensive, though I haven't even found a self-contained one that's notably below 500 g.
I'd like to know the polygon count, both vertices and faces, if possible.
I'll check the site and write more tomorrow, with the sun coming up shortly I should catch some sleep.
Also, yeah, rolling shutter can be compensated, but the impact goes through all the pipeline steps. All of them.
I don't want to re-write the whole pipeline. Worst case I'll settle for an Apertus Axiom Beta.
Also, a global shutter works nicely with a strobo, and a NIR-enhanced sensor works well with e.g. 808 nm light sources. The reason would be around power conservation when the robot has to bring it's own light source.
My last experiment had about 5k or so depth maps getting reconstructed (15 fps 4k Mavic Pro footage, iirc 100Mbit/s), from barely over 20k captured frames that I did full SfM on. The reconstruction density was adaptive, manually choosen to ensure sufficient scene coverage without spending too long on the depth map reconstruction.
The issue is that you have to assume somewhere between 1/60 and 1/500 exposure time for motion blur, if operating without an electronic global shutter. And I like detail, so I'm more looking for something like a 50 mm lens in front of an academy-sized sensor, yielding about 70~90 mm 35 mm-equivalent for the typical 4k by 2~3k sensor.
Reading your exchange I get the impression that a production camera would be suitable for photogrammetry?
Many older (and now inexpensive) Blackmagic models have a global shutter, and shoot RAW at 25 fps / 4k or higher.
Would you reccomend a wide angle lens for this purpose? Intuition tells me a telephoto would capture more detail, though with the caveat of requiring more movement over the surface area to be captured.
The first variable is what you want to capture. I am trying to capture large outdoor scenes in suitable detail to be used for robotics simulation. A telephoto would not work. I am preferring very wide lenses or even fish eye.
But some people are capturing individual rocks to be used for video game assets, or surveying historical sites or something else. There are times when a telephoto would be desired, especially if you cannot get close to the target scene.
4k is 8.5 megapixels so it’s relatively low detail compared to my 18 megapixel video from my Panasonic. I personally do not really know the math behind any of this so I am not sure where the trade offs are between high resolution rolling shutter or lower resolution global shutter.
Because I am doing scene capture for robot simulation, I do not need a perfectly accurate model. I would rather have a machine learning algorithm that places simple polygons where real world objects are in the video and gives them a reasonable texture. That is, I want to go from video directly to a low poly textured 3D model. My problem is complicated by the fact that I am developing an off road robot, and photogrammetry of forests is very hard.
This is why I mentioned machine learning. I care a great deal about the semantics of my scene and less about metric accuracy. What kills me is compute time. Finding feature points in every image and matching them all together is extremely compute intensive. I believe shortcuts can be discovered for problems like mine with creative neural networks. That said, I still need to learn more about existing algorithms.
Sure, if you get global shutter and resolution equivalent to about 5 MP bayer sensor or 2 MP monochrome sensor, and get output with an intra-only codec, that's fine.
Those black magic cameras are severely bandwidth bound, compare to e.g. the Pocket Cinema 4k (USB 3 and able to fully load a USB-C SSD), which unfortunately only offers a rolling shutter.
The lens is only slightly relevant, in that an extremely long focal length will cause issues in estimating said focal length, and typically also in depth-of-field if that's relevant for your scene. I suggest no ultra-wide-angle lens and no more than 500 mm (35mm-equivalent). 30-150 mm focal length (35mm-eu) would be the range you want to look at.
I wish monitors in general had better (and open) firmware. It would be really cool if for example I could have my laptop plugged into my desktop's monitor and have it show up as a "window" on my desktop. All it would take would be for some simple driver on my desktop to tell the monitor where to draw the image when I moved the window around. Basically a better version of the mostly useless PiP feature my monitor already has.
In this particular case, no, as the image wouldn't be being sent into the PC to render in its framebuffer; but rather the PC would just be drawing an empty window, and reporting the geometry of that window to the monitor. The monitor, with two HDMI leads plugged in, would be responsible for compositing the inputs together, according to the geometry the PC reported, but all internal to itself.
That's how hardware-accelerated video decoding used to work in Windows XP days IIRC (before GPU-based desktop composition), the video player would be a blank black square and the GPU would be told to draw the video on those coordinates.
Because of how it was implemented, you could drag VLC around while the video was playing and the video would stay "behind" everything, with the VLC window acting as a "hole" through which you could see it. (So you could move the window to the left and see half a black square on the left, and the left-most half of the video on the right)
Nowadays with desktop composition AKA DWM, Windows just makes sure to black out DRM content from any frames/video it sends to an app requesting to capture the screen, making sure to send the video-including composed desktop only to the display. (And if you have some GPU recording software like NVIDIA ShadowPlay, it switches off when DRM content starts playing) You can see it in action with the Netflix UWP app. Of course, a bunch of DRM-respecting software -- like Netflix in Google Chrome -- doesn't really follow that spec and can still be screenshot/video captured like any app.
It provides lower resolution content to those devices, capping at 720p.
Which is the same as Netflix effectively, Netflix just provides higher to PCs when they can block screen recording entirely. Disney might eventually I guess.
This isn't a firmware issue, because enabling this would require adding the hardware to the scaler ASIC to actually process multiple video streams, and to increase the buffer size and bandwidth n-fold so that it can synchronize and composite the unsynced video sources (also introducing 1+ frames of latency).
The GP was assuming a scaler that already supported a Picture-in-Picture or split-screen feature—as many modern monitors do!
The problem—and it is purely a firmware problem—is that all such monitors make this kind of support "dumb", with hardcoded geometry (i.e. split-screen as exactly halving or quartering the screen; PiP as putting one input in an box that composites over exactly the upper-left quarter of the lower-right quadrant of the screen.)
There's nothing in the scaler ASIC that particularly benefits from these numbers being hardcoded, such that its job would be any more complex if they were controllable via machine-registers POKEable using HDMI-CEC commands.
Monitors are generally connected through DisplayPort or Thunderbolt, not HDMI, so there's no reason for HDCP to enter here. The card in question does have a HDCP output, but advanced features are only available through DisplayPort.
It might surprise you that most monitors out there, especially the cheap garbage ones from Amazon/Walmart that most consumers have in their homes, are connected via HDMI.
As someone running their monitor over HDMI, this is news to me. If it gives the resolution and framerate of the monitor, what's the difference if it's HDMI or DisplayPort?
Displayport used to be much better but now they both cover virtually everyones needs. Now the main difference is HDMI has a licensing fee. Some major TV companies are in the org so they don't have to pay it which is why TVs always use HDMI and GPUs usually have more displayport because those companies have to pay for hdmi.
For you as an end consumer nothing if bandwidth is enough with either cabke, but techies like us will be sticklers that one is more open than the other, not bound by royalties and such.
That, and some video cards support higher resolutions over DisplayPort. For example, while my 2013 Mac Pro has both HDMI and Thuderbolt/Mini DisplayPort outputs, the only way to get 2160p60 HDMI output (without chroma subsampling) is via a DisplayPort-to-HDMI 2.x adapter (and I had to try about half a dozen different adapters before I found one that does this correctly).
Displayport usually has something like 50 % higher bandwidth compared to the HDMI standard of the same vintage. Which makes sense, since home video folks are fine with like 15 pictures per second and 4:0:0 chroma subsampling, but that doesn't really work for computers...
> You plug your mouse into the USB ports on the monitor and it tells you how long it takes the image to change after you move the mouse. Brilliant idea.
Yes. The industry has been optimising for throughput in the past decades. It is time to bring latency back to the table. I want super low latency Computing. Street Fighter or The King of Fighters in CRT Arcade Era just felt so much more responsive.
Not at this time. There are some hypothetical options like the LG 27GN950, but it has poor contrast and uses Displayport 1.4 compression, which is only supported in the latest graphics cards. VESA and friends really made DisplayPort 1.4 as close as they possibly could to false advertising without actually committing it (because the only relevant new thing in DP 1.4 -- DSC, display stream compression -- is technically optional and no one claimed they'd support DSC while saying they support DP 1.4, which is pretty much the same as supporting DP 1.3, since nothing of note changed).
And now we're looking at DisplayPort 2.0 which can already barely support uncompressed 8K at 60 Hz and is basically maxed out by 5K/6K 120/144 Hz. And it's unclear if the presently introduced generation of GPUs even supports it, or if we're going to use effectively-2013-Displayport until ~2022.
Note how the marketing material only talks about HDMI 2.1; DisplayPort isn't mentioned once.
I think DisplayPort is not mentioned because Nvidia hasn't upgraded their cards to 2.0. Super disappointing. 4k/144Hz/HDR is enough for now, to be honest, but DP2 can do 84/144Hz/HDR as long as it has DSC, 244Hz with subsampling. The slow pace of development and deployment has hurt everyone for sure, but I don't think the standard is bad in itself - just a bit late.
Yeah, good points–it seems even armed with a 3090 we'll be limited to 8K/120hz/12 bit over HDMI 2.1...we'll have to wait a while before we're actually driving 8K at higher refresh rates.
OLED TVs are in a much better space than PC LCDs, where no really good options exist. Either IPS, which has poor contrast, bleed and doesn't do HDR, or VA, which has good contrast, but also doesn't really do HDR, and VA generally has poor uniformity (nitpick) and viewing angles. Some VA are pretty smeary, but that seems to have cleaned up in the latest generation. TN panels are much better than they used to be in the color department, but it doesn't have the contrast of VA, and even poorer viewing angles than VA.
OLED is clearly the way forward - accurate colors, excellent contrast, no bleeding, no uniformity issues, proper HDR, excellent response time. Except OLED doesn't come to PCs.
Some things could be counter acted with software (make desktop taskbar only visible when mouse reaches bottom of screen, etc.), but I would be interested in some real life data as to how long it would take for even burn out over the whole screen, till it drops as much, that the quality degrades to TFT level.
I'd be okay buying a new OLED monitor every year if they were available in the format I want (large ultrawide) and with low latency scalers, high refresh reates and without privacy destroying "smart" features plaguing TVs. Anything to get out of the crappy tradeoffs there are with backlit displays.
Iiyama's MVA3 displays are, to my knowledge, significantly in front of other LCD styles (e.g. TN, IPS) as far as contrast, especially from an angle, is concerned.
Unfortunately my desk's dear centerpiece, https://iiyama.com/gl_en/products/prolite-x4071uhsu-b1/ , has been EOL'd about 2 years ago, because I have been unable to find a replacement that's not worse, while staying in the 40~50" range. Any concrete suggestion would be greatly appreciated.
Its not really full end to end latency though. I think, at best it can measure when the the monitor thinks it received the signal from the mouse and when it perceives the new image. Latency in the monitor itself can't be measured, can it?
Is there any reason to believe latency within the monitor itself wouldn't be relatively constant? Or would you expect large variances in monitor latency? If it was a pretty narrow distribution they could simply add on that constant factor. I'm not sure if they do this though.
I recall John Carmack being interested in measuring latency in exactly this way during his Occulus tenure (think he mentioned it in one of his keynotes). I'm sure he would be pleased to see this is now becoming mainstream :)
In addition to input latency, with "real time" computer graphics it's basically impossible to know when your frame will actually be rendered on screen. If this information was available, it would be a lot easier to achieve smooth visuals without having to chase higher and higher frame-rates.
Don’t GPU fences tell you which vsync you’re targeting? Or are you saying you want the actual estimated display time of the frame? I know the latter is definitely available for VR systems as it’s required. Having it available for 2d games probably would help marginally in cutting down the perceived latency although I don’t know if you’d be able to feel it.
John Carmack has mentioned on Twitter that he spent a lot of time trying to make the frame release rate more consistent even with purpose-built VR hardware. I think game programmers found it easier to go with the gsync/freesync method of doing a vsync whenever the frame is ready instead of trying to hit a deadline.
Wow, I wasn't planning on upgrading from a 1080 (non-Ti) but the 3080 is so good and priced so well, I probably will. I just ordered a 240 Hz 1440p IPS monitor and I wasn't planning to hit 240 Hz, but this makes it so easy I might as well.
My day job is primarily ML as well, so I might just go for the 3090. 24 GB of memory is a game changer for what I can do locally. I really just wish Nvidia would get its shit together with Linux drivers. Ubuntu has done some great work making things easier and just work, but having them directly in the kernel would be so much nicer.
One thing I'm curious about is the RTX IO feature. The slide said it supports DirectStorage for Windows, but is there an equivalent to this for Linux? I'm hoping someone with a little more insight or an Nvidia employee may have some more information.
> I really just wish Nvidia would get its shit together with Linux drivers.
This has always confused me. They are pushing ML hard, yet we often use Linux for that kind of work. And these cards are the ones expected to be used in universities and home labs. Linux drivers that "just work" would be a big push forward to really show that they are trying to push for ML development.
I have several nvidia GPUs of different generations on Linux, and Nvidia drivers + CUDA "just work". They are closed source - that's bad - but I don't get what people are complaining about related to "just working". They DO "just work" (Arch Linux, both package and BLOB install...
- Ubuntu 18 I had hit or miss with the hdmi connection on my laptop (1060Q). 20% of the time it would work if I plugged it in. 50% of the time it would work if I had it plugged in and rebooted. This makes giving presentations difficult.
- Arch/Manjaro/Fedora/Ubuntu 16 I could never get the hdmi connection working for a laptop.
- All distros, difficulty getting cuda running AND using the display. Intel drivers for display + nvidia for cuda works, but this means I can't use my GPU when I want to do some of the limited linux gaming.
Laptops seem to have more problems than desktops. On desktops I have many more hits than misses (if I have the graphics card installed when I install the OS). Laptops have just been terrible.
The headache with Nvidia drivers and Linux is that depending on how new your model is and what distro you're using, you may or may not be able to load the GUI until you've manually upgraded kernel and disabled the Nouveau drivers (if enabled) then updated to latest nvidia-drivers-XXX package.
I tried to get Ubuntu/Pop_OS! 20.04 running on a dual-GPU laptop that has an RTX-2060 and AMD Renoir integrated. Had to modify kernel boot params to disable nouveau modeset and then run script for mainline kernel upgrade + drivers to get it to run. Was not a fun Saturday =/
But prior to this, yeah Nvidia non-open source drivers have mostly "just worked" for me on the older models.
Most laptops with any Nvidia hardware have two GPUs, one less power-hungry GPU integrated into the CPU and one dedicated GPU to play games on when you're connected to/sitting nearby a power outlet.
As good as the performance of mobile GPUs is these days, they still suck up a lot of power. I don't think you'll get much work done on the go with a laptop that only has a dedicated AMD/Nvidia GPU.
Normally the low power gpu is the integrated Intel one that's on all their consumer chips, so it's more adding one gpu than slapping in competing ones. Although it does cause driver issues anyway, hence the thread.
Something interesting I've noticed is that there is a big difference in usability between desktops and laptops. Desktops I have far less issues. But laptops I have to get lucky to be able to use my HDMI port. And good luck doing that and having cuda support.
Out of curiosity, what laptop models + distros have you tried to set up?
On Acer Nitro 5 with Nvidia GTX-1060Ti and i5, stock Ubuntu 20.04 loaded no problems, and even DisplayLink driver for dual monitor, one through regular HDMI + other through USB -> HDMI adapter worked (though I couldn't get it to rotate display vertically).
The bugs I did have were with it constantly re-disovering network printers that I had to disable, and changing the default wifi power-saving settings because something was funky with it.
On Asus TUF A15 with Nvidia RTX-2060/AMD Renoir + AMD Ryzen 7 4800h absolutely no distro worked out-of-the-box and I needed mainline kernel + latest Nvidia drivers. But after fixing that myself Pop_OS picks up everything perfectly and no problems.
Both of them have CUDA working IIRC (at least running "nvidia-smi" says it does).
I have an inspiron with a 1060Q and had similar problems on an HP envy (forgot the card in there). I've tried a bunch of distros (I've been on linux for about a decade and mainly run Arch though).
I have heard great things about Pop and I am going to be building a new machine with these new cards and giving pop a try.
> except more driver patches and performance tweaks
Honestly, to me that's not a big sell lol. And probably not to the type of people that like distros like Arch. We just want nivida to work like AMD and mesa drivers do. (Having done some testing, I am able to squeeze more performance out of a distro that builds up, such as Arch, vs a distro that you take down, like Ubuntu).
This is right. I am running Pop OS, but will make the jump to Arch when I have time. I don't want to wait for the distro maintainers to release new kernels/language versions/etc. And once you start changing those things, the benefit of the distro starts to fade away.
A nice middle ground is Manjaro (there's a nice GUI to select your kernel). Though I'm not sure what nvidia settings Pop uses, because I still have the above issues with Manjaro. But this may again be laptop centric.
I've used Manjaro on two legacy laptops and a fairly recent desktop and I can only vote for it for almost anything you would want to do with your machine. It's a very lightweight and well-supported distro that has outperformed machines that are on paper twice the speed of the one it's installed on.
Not sure how easy it is to actually game on it though, haven't tried yet. For everything else however, Manjaro is an extremely solid choice, with a very recent kernel too.
The Linux drivers refuse to properly detect the supported resolutions on any of my monitors. They get the native resolution, plus 1024x768.
This is real frustrating when I want to play older games that only support 4:3 and 5:4 resolutions, because I have to knock them all the way down to 1024x768.
There's a roundabout way to manually add custom resolutions, but I've never gotten it to work with X. Maybe it would play nicer with Wayland, if Nvidia made any real effort to actually support it.
Hell, even Canonical tried to sabotage it by starting their own NIHing alternative but thankfully Intel told them promptly to fuck off if they expect them to support it.
That doesn’t really answer my question.... rephrased: all I do with high-powered GPU’s is Remote-SSH into hosted Linux machines and run ML jobs. Does this wayland thing have any impact there?
It should answer the question because the use case which you are stating (didn't before) is that you're using a headless environment. Which means you aren't using a DE.
IF you are using a server with a DE then it answers the question about how yes it does affect you because you are using a DE.
NVidia has always been a headache for me. I run ATI on Linux. I use xmonad as a window manager, and NVidia's proprietary drivers just didn't handle multiple screens correctly. I switched to free drivers, and they did one fewer monitor than I was driving. I switched to ATI, and then things just worked.
What's more surprising is that ATI doesn't take this opening.
At some point Nvidia crippled their drivers if you didn't have a Quadro on Linux. Eg: No multi-monitor. Not sure how of that they do now.
So some guy resoldered some resistors on a GTX 690 (at the time the highest spec card they sold). To make it identify as the Quado with same specs.
"Just works" before ticking a "Install privative driver", and then getting many bugs.
Compared to AMD drivers that just works without additional installation, have far less errors and play well with the standards.
The igpus on the 10th gen intel desktop chips I've got have been working well without any hassles, I'm running manjaro kde (and previously kde neon) on a 10400 and a 9900k igpus in wayland, and they work great!
i think the WM i'm using is kde plasma that comes with manjaro currently, i set the compositor to use opengl 3.1 using wayland from the plasma-session-wayland package
i'm using google chrome, vscode insiders and alacritty and vlc, and that's pretty much it, with very little problems, after i boot into wayland and launch chrome i get some stuttering on the shadows which goes away after a couple of seconds, but that's pretty much the only hassle
My experience with a 1060 on Ubuntu is that it was fairly buggy in the early days, and has gotten better.
I don't have too many problems now, but just recently I ran into trouble with any application that used OpenGL not working, and it went away when I rebooted.
In general I'd say they've done an okay job, but still I've seen enough bugs to say they have room for improvement.
I don't know how many of those bugs are Nvidia's fault explicitly; Linux is a bit of a wild west when it comes to software configuration, and they can't test everything or fix every buggy application.
Because some DEs like KDE have devs behind them that don't want to deal with non Open Source blobs out of principle so you get a terrible experience with Nvidia. Gnome is usually fine.
What package do you use for arch? Are you able to run wayland? What windowing system do you use?
I've just switched to manjaro kde and I'm having a blast using it every day for work, any tips on the nvidia drivers is appreciated. I have managed to brick my distro a few times trying to get it set up (and by brick, I mean getting it into a state that I can't figure out how to recover it from).
I've had an M4000 in my desktop and a P620 in a side-box, both running SuSE Tumbleweed for a year or so using the NVIDIA drivers. Once I got the initial setup done (follow the directions), I've had zero issues. Updates have been smooth, though there's an extra step now (hit ENTER to accept the NVIDIA license). GLMark2 score is a bit north of 7000.
They are not part of mainline linux, so if you do install newer versions, there would be broken changes that you wait for Nvidia to fix or have to fix yourself.
I have the same confusion, although I will say the only time I've EVER had a problem with nvidia drivers it was when I installed it through the package manager.
When I install them myself by running the scripts that nvidia gives you, I've never had a problem with the drivers. I've always suspected that was the problem, the way the distro's package the drivers, rather than being a problem with the drivers themselves.
I wonder how many expensive gaming cards Nvidia has sold to programmers saying "Well, I can't really justify spending this much to play video games, but what if I want to train a DNN at home?"
> One thing I'm curious about is the RTX IO feature. The slide said it supports DirectStorage for Windows, but is there an equivalent to this for Linux? I'm hoping someone with a little more insight or an Nvidia employee may have some more information.
I haven't been able to find any real technical details about what DirectStorage really is, but my expectations are that it will consist of:
1. An asynchronous IO API allowing for low-overhead submission of IO requests from the application to the kernel, including batch submissions and reaping of completion events.
2. Some provision for transparent compression offload, either to the GPU or to a CPU-based fallback
3. Optional support for peer-to-peer DMA transfers of data from SSD to the GPU's VRAM
Linux is already the gold standard for (1) with the relatively recent io_uring API, and has support for (3) to some extent (P2P DMA has been a fairly obscure feature until recently).
There are still some pretty big unanswered questions about DirectStorage. How well will it reduce the IO overhead that currently allows antivirus and other programs to hook into the IO stack? Will it be compatible with non-Microsoft NVMe drivers, including Intel's RST drivers that are commonly used for their software RAID? Microsoft doesn't seem to want to make that kind of information public this year.
Windows NT 3.5 (1994) would beg to differ with its support for IO completion ports (IOCP). I think io_uring is more general and more flexible for the sort of M:N scheduling systems now commonly used in programming languages, but IOCP predates io_uring by 25 years!
The rest of your comment sounds right to me, there are unanswered questions about how DirectStorage interacts with filesystem filter drivers.
Being old isn't particularly interesting here. io_uring can do 2.5M IOPS per CPU core with a sufficiently fast SSD. I've never seen any suggestion that IO overhead on Windows can be cut down enough to get close to that. And if IOCP was actually good for this use case, Microsoft wouldn't be hyping a new DirectStorage API. (It might turn out that DirectStorage is new window dressing on top of IOCP, but I suspect they've had to cut stuff out of the existing IO stack to reduce overhead.)
I think my point was largely that Windows already has (1) from your list and it's already the gold standard on Windows, and has been available for some 25 years. I would be surprised if it wasn't part of the mechanism used in DirectStorage.
io_uring does have a more ergonomic and POSIX API, but that's because it's on Linux! Fundamentally they're both completion port based replacements to polling for IO events and they're both capable of high throughput and horizontal scaling.
Ok, the misunderstanding here may be that I was using "gold standard" as a superlative among async IO APIs in general, while you're using it as a superlative among a given platform's set of IO APIs, async or otherwise.
I was extremely (I think justifiably) skeptical of async I/O on Linux for the longest time but now I legitimately think with io_uring Linux is actually somehow the gold standard for (1). They really got their act together. io_uring is really good.
It's the same tech used by the next-gen consoles (e.g.: PS5).
Essentially, it's about streaming game textures directly into the GPU memory.
The idea is that this is highly asynchronous and parallel from the perspective of the CPU, so trying to do this through the "main game loop" would be an absolute nightmare to code at the scale we're talking about here. Think 50K IOPS sustained at 8GB/s with pretty hard real time guarantees.
Yeah, I already read that. If you pay attention, you'll notice it talks a lot about the problem they're trying to solve, and very little about how exactly they're going about solving it. Partly because it's apparently still in flux, which calls into question what's going on with the storage API for the Xbox, which really should have been nailed down by now.
Now that the reference cards don't have the blower coolers they are a lot more tempting. I'd still wait for reviews to see how loud they are though.
Up to now, OEM/AIC/AIB cards had quieter cooling systems, but they also make you pay for the silly graphics, the RGB, etc. I don't think the overclocking ability is really that much better. A quiet reference design with a non-blower cooler and a good warranty is what I've always wanted.
I have a Micro-ATX case for my PC and am a little skeptical that flow-through would well for that form factor because the intakes are on the bottom of the card.
It will probably be fine. I've used powerful graphics cards in super-cramped ITX cases and didn't have a problem. I think people overestimate how much cooling these things really need.
Any temps below 80C at load are perfectly safe and shouldn't result in throttling -- the same is true for CPUs. If you aim for lower temps than 80C, you aren't going to get better performance, although lower temps are a good way to get your computer to run more quietly.
I know for the 20 series Nvidia supplies AIB's like EVGA with binned chips too, higher end cards had GPU's with slightly different model numbers that typically clocked better. They probably do the same for all the different card makers.
Basically, after the chips are manufactured, they're not 100% uniform. Some have better performance, some have worse.
In this case, it means they're reserving the best chips for the founders cards. In other cases, there have been instances where a company has two products, a high end and a low end (or medium, etc). In some of those cases, people have investigated and the chips are actually exactly the same, but the lower end product will have a core disabled or similar, depending on the exact product. That'll happen a lot of the time when the company has yield issues where too many of the chips don't have acceptable performance or one part of the chip is just broken. They'll disable the broken portion and boom, the lower end product is born. That's still a net win for them because the alternate is either to throw the entire thing away or spend more time improving the yield.
chip manufacturing is an imperfect process, and so there is variance in performance/viability of all of the hardware on a chip. the higher performing chips are "binned" for the top end of the price point, while the lower performing chips are either binned for lower performance or have some of their functionality disabled. For contrived example, A company may produce nothing but quad core chips but sell those with some cores that don't meet minimum performance as dual cores with the bad cores physically disabled.
they cherry-pick their best silicon so you can run them at higher speeds if you're overclocking or lower voltage if you want cooler temperatures and less power consumption at stock speeds.
There's always a ton of speculation about it, I've seen the claim NVIDIA bins for FE cards and factory overclocked AIB cards together
At the end of the day it doesn't really matter, you're paying a FE premium for early access mostly.
On the plus side, this time the FE card might have a top tier cooling solution, which is why I'll probably be caving to their FE tax (and probably plenty of others, focusing on cooling was a smart move)
I don’t believe it, old reference fans were optimized for static pressure on server chassis and gaming cards optimize for low speed high volume airflow, I think.
The terminology is a bit confusing: OEM in this context is referring to aftermarket (NVIDIA partners) cards. It's referred to as "OEM" because the aftermarket cards come from OEMs like Asus, EVGA, etc.
A rough inflation guess is 330 six years ago is 363 today. The 3060, which is still a relative monster, will set you back 400.
My last gaming computer took me 9 years along with only an SSD addition, RAM upgrade, and video card change halfway through. It was only Microsoft Flight Simulator that made me jump up and damn... my timing was a just a little too early on the beefy video card.
Am in the same situation, the 1080 is still holding up well, but the 3080 looks very tempting indeed.
Looking forward to some benchmarks that evaluate the difference between PCIe 3.0 and 4.0. Getting a new motherboard with PCIe 4.0 along with a CPU will up the cost considerably, so hopefully that can be avoided.
I use an eGPU, so I have a 4x instead of a 16x PCIe 3.0 link with a 1080Ti and even that doesn't make much difference, some games blit/do more system <-> gpu transfers than others (APEX, I'm lookin at you) and there's a noticeable perf impact for the most part it's negligible. I wouldn't worry at all about PCIe 3.0 vs 4.0.
Same. In gaming and graphics benchmarks on my 1070Ti over Thunderbolt 3 (Hades Canyon Intel NUC), I've seen no more than a 10%-20% drop in frame rate vs. what I'd expect from equivalent benchmark results reported online, so I'd expect no material difference between PCIe 3.0 x16 and PCIe 4.0 x16 for typical gaming applications, and, from benchmark numbers I've seen, only a modest performance drop from PCIe 4.0 x16 to PCIe 3.0 x8.
This assumes, of course, that the PCIe lanes in question actually exist† and are dedicated to the card, and are not shared with another in-use device via a switch or similar chipset shenanigans.
Case in point: the only time I ever had performance problems with my eGPU setup was when attempting to do 4K HDMI output through a Blackmagic Decklink card while simultaneously using the 1070Ti for compute over the same (shared) Thunderbolt bus in DaVinci Resolve.
As far as I know, this would not typically be a problem with a single-GPU, non-Thunderbolt-connected desktop system, as desktop motherboards typically have at least one full-bandwidth x16 slot, which you'd almost always use for the GPU.
† Physical x16 slots can have fewer than sixteen connected PCIe lanes. For example, while my HP Z820 has four "x16" slots, only three have full bandwidth; the fourth slot supports x16 cards, but only at the speed of an x8 slot, because only eight of the slot's sixteen PCIe lanes are actually connected.
You can find the information from multiple sources - IT review sites that performed the testing - that for current GPUs PCIe 4 x16 is not needed, you can use PCIe 3 x16 with negligible impact and even PCIe x8 with minimal impact.
Great to hear that PCIe 4 is not a necessity at this point.
About the PSU, for me the only reason to go with a 1000W PSU is that the fan will remain quiet with a GTX 1080 when idle and gaming as well. At least thats what I hope.
If the same is possible with the 3080 is more questionable, especially if the card draws 320W at all times.
That capability exists even on smaller power supplier, for example the Corsair RM-650 has it; powering the fan is done based on temperature (and that depends a lot on the load), not on the power supply max rating. This means you need to just add your max CPU consumption (not TDP, that is lower), GPU max consumption, the rest of the components and add a 200W reserve, not a 500W reserve.
PSU efficiency is low at the ends of the curve, underutilized (<20%) the efficiency can drop as low as 70%. The top (not peak) efficiency is somewhere in the 30-80% load zone, that is where you want to be.
The major difference is offloading some of the load from the CPU when using the pci-e bus.
If you have a spare core not being used when playing a video game, it can be used which should make the difference between 3.0 and 4.0 somewhat indistinguishable.
>> My day job is primarily ML as well, so I might just go for the 3090. 24 GB of memory is a game changer for what I can do locally. I really just wish Nvidia would get its shit together with Linux drivers.
We have a small team of 7, so after a lot of bureaucratic navigation I got us custom built desktops for local experimentation with 2080 Ti's and for the rest we use your typical cloud providers. Personally, I'm partial to GCP, because of the TPU support and BigQuery, but AWS is fine too.
> One thing I'm curious about is the RTX IO feature.
Me too, but for a different reason: I wonder if it can work with BitLocker if you're using software encryption (as hardware encryption is transparent enough that the drive basically locks and unlocks).
It would still save the CPU cost of decompression but it'd have to go through RAM for (AES-NI accelerated) CPU decryption either way. Maybe at that point RAM speed and latencies start to matter more. Or the feature turns off altogether. Definitely something to test.
Yeah, Wayland is one of the main issue here. Things are going really well generally speaking with the Wayland transition, but Nvidia is single-handedly delaying the wide-spread transition from X to Wayland across most distros, which is super unfortunate, because it's a very important part of a desktop Linux experience. Using Wayland day to day on my XPS 13 with regular Intel integrated graphics is great in terms of perf and battery life. There's a much larger conversation to be had about X vs Wayland, but broadly speaking Wayland is just better and not bogged down by design decisions made 20 or 30 years ago.
The other key benefit would be driver management. Updating your drivers right now can be a nightmare. If the drivers are open sourced and upstreamed into the kernel this becomes a non-issue. I do understand why Nvidia doesn't want to do that though. A lot of their lead right now is not just hardware based, but software based. They have a choke-hold on the ML ecosystem and it's a huge cash cow for them. Giving away that secret sauce in their drivers, so that AMD could make their cards seamlessly compatible would probably be a huge mistake from a business perspective for Nvidia.
> Giving away that secret sauce in their drivers, so that AMD could make their cards seamlessly compatible would probably be a huge mistake from a business perspective for Nvidia.
I've seen this argument a lot but never from the ML perspective - and I don't think it makes sense with ML.
Nivida's advantage is ML is CUDA and cuDNN, and the huge set of tools built on them. These aren't the driver, but a layer above it.
I don't really understand the "secret sauce" argument for the drivers at all, but I assumed it applied to gaming somehow. I can't think of how it applies to ML though.
They support nouveau. Wayland doesn't support EGS, no matter what DE you stick in front, so nv drivers are out.
(Nvidia cards have two drivers on Linux, nouveau, the open-source, slow, and incompatible-with-Quadro ones, and NV, the binary blob shipped by Nvidia that people take religious exception to.)
Gnome supports Wayland over EGS just not by default. I tried it out a few months ago and while it "works" it was somewhat buggy and I just went back to X.
> My day job is primarily ML as well, so I might just go for the 3090. 24 GB of memory is a game changer for what I can do locally. I really just wish Nvidia would get its shit together with Linux drivers.
You might be better off waiting for Quadro.
GeForce, from NVIDIA's perspective, is the consumer line. The drivers are focused on getting the best possible performance for gamers (on Windows you'll see a big deal made of "Game-ready" drivers) and the hardware is intended for desktop usage, i.e. a couple of hours of gaming. GeForce hardware isn't intended for long-running jobs like ML and the drivers aren't stability-focused.
That's not to say that you can't train ML models on a GeForce card or that the GeForce drivers will lead to constant crashing or anything, just that NVIDIA isn't focused on this for GeForce.
Quadro and Tesla on the other hand are all about reliability. They're intended for use in servers and workstations where they're subject to all-day (Quadro) or 24/7 (Tesla) load. The drivers are focused on making sure things don't break.
Though I think the 3090 is intended to be the successor to the previous-generation Titan, so it could be that NVIDIA has different ideas about what goes where these days.
There are indeed driver differences but I have been running long-lived, sometimes month-lived ML workloads on GTX and RTX cards for years. I've never had any stability issues.
You are more or less buying into their marketing, and paying 100-250% price premiums for a "Quadro" or "Tesla".
> the hardware is intended for desktop usage, i.e. a couple of hours of gaming.
Tons of people have been flooring previous generation of cards 24/7 for months with compute workloads, and not nearly all of them were using watercooling. Those cards are still fine.
ECC memory might be an argument though, for CAD/CAE type of work. Doubt ML cares about a few bitflips.
Tons of researchers and companies use the GeForce line for ML. It’s the reason for the infamous “no use in data centers” addition to the EULA, and potentially part of the reason for the switch from blowers, which made it feasible to stuff 4-16 1080Ti’s in a single case.
Having gone through some hellish months on a 5700xt, AMD isn't any better than Nvidia on the Linux front, though since late January it was relatively stable (on beta linux kernel, which meant I couldn't use KVM/Virtualbox reasonably).
Personally, looking at the RTX 3070, for 1440p, but I'm not doing anything other than some casual gaming.
Just so you know, AMD is also releasing a GPU update this fall. Could be that drivers take a moment to stabilize again, but at least the old models should drop in price if you're not looking to be on bleeding edge.
Yeah, I'm hoping they're more stable out of the gate... from October through early January the navi cards were problematic to say the least.
Kind of works in kernel, update breaks it, switch distros to supported distro, load official driver package, update breaks it... switch to beta/bleeding kernel, graphics work, but kvm/virtualbox completely borked. I may do a clean install with PopOS 20.10 when it drops, which should be well supported in kernel and get virtualization support back in the box.
I'd give it a few weeks, should be able to get one used around $200 after the rtx 3000 availability shakes out. A GTX 1080 is around the same performance and pricing will be a little lower even.
Since January it's been pretty stable, I've been on the beta kernel releases though.. when say 20.10 ubuntu and derivatives drop, should be using a kernel that's stable for the 5700xt.
I've been working more on my Windows drive, using WSL2 + Docker Desktop (wsl2 option) for the past few months though, which I've been surprisingly happy with.
You have to be running Windows 10 version 2004 update or newer. There are lots of walkthroughs, but you install "Ubuntu" or another distro for under WSL via the Windows store.
If you install docker desktop with WSL2 installed, there's a configuration setting to "use wsl2" it may still be labelled beta as a feature. But it does work better.
I'm using the new Windows Terminal, and my default is set to WSL2 Ubuntu... for all intent and purpose, I launch VS Code from a wsl2 prompt/directory and it's connected into "linux" ... my terminal and code terminal/directory are in linux, and I use it like any linux distro from that pov. The desktop is windows, but almost everything I work in is via Linux under WSL2.
The G9 and G7 seem to suffer from the same problem that previous VA adaptive sync monitors had, i.e. brightness depends on frame rate, so the screen flickers with changing framerate.
In addition to that they seem to have some issues with their firmware trampling over its own memory, causing glitch artifacts and such. I suspect they'll fix that through a firmware update.
The EVE Spectrum[1]. I actually pre-ordered a while back, so got it for less than what it's priced at now. Their first project had some issues with delivery, but they seem to have done a great job with this one and I'm cautiously optimistic.
I wish they'd get their drivers sorted out too! I'm typing on a computer with a 2080 ti, but instead I'm using the igpu for kde wayland (which works perfectly fine).
A small thing that I haven't seen mentioned yet, but these cards have native hardware decoding for AV1. With chrome just launching support for AVIF this last week it seems like more and more platforms are getting out of the box support for it. Nvidia is also working with a lot of partners[1] on it it seems. I'm currently working on a social video platform, and having a unified next-gen codec that works everywhere for both images and video would be SO helpful. Hopefully this trend progresses - would love to be able to do some extremely high quality streaming.
There’s not a ton of money in that since most people who would need fast encoding are streaming, and those folks already used capture cards or secondary PCs to record.
It would be super useful if it could be done on mobile SoCs. Realtime AV1 encoding on mobile phones would be huge. But even if you don't care about speed, AV1 encoding is still painfully slow. Encoding an hour of video can take over a day on a modern cpu.
It's interesting the effect new graphics have on how old graphics are seen. I remember when Resident Evil came out on the first Playstation, at the time, I considered certain game elements like fire and water indistinguishable from reality. Now it looks like pixelated garbage and I ask myself how I could have ever thought that.
Part of it is also that modern TVs and monitors are much higher resolution than their predecessors.
One interesting point of reference is the initial switch to HD in the 2000s; I remember there was a bit of panic in the beginning because news studios had to adjust studio lighting and makeup; flaws that were not perceptible on a CRT were all of a sudden extremely noticeable blown up on a 48" flatscreen.
IMO watching a few sci-fi movies in 4k+ looks ridiculous - I start noticing the difference between CGI environment and actors and it kills the immersion completely, it goes from "that character is really there" to "this guy is larping in front of a green screen"
Higher resolution, better color replication, and frame rate make very obvious the fact that there seems to be a magical glowing orb following around the characters right behind the camera. Immersion breaking because you can get away with it with less quality, it's more difficult to notice.
Something that I've also found more and more irritating is the foley artists doing ridiculous things for sound effects, especially in nature documentaries, but all over the place really.
that's not the point - they were saying that retroactively going back and viewing old content on 4k, you can easily spot the CG/issues etc.
ex: Farscape - all the CG was rendered in low res, so even when you view a 1080p copy it looks silly. Imagine all the content that would have to be re-done and upscaled to make it watchable in 4k.
Based on anecdotes it seems like it. Older generations always seemed to think pets had no idea what was going on when viewing a TV to the point of ignoring it. I'm assuming their body language suggested so.
Nowadays people open up koi pond videos on phones and let their cats play with it. But like you mentioned, it would be interesting to see a study on it.
The framerate didn't increase. It was still 60 or 50 Hz. The resolution did increase, but so did the size. The typical angular resolution stayed pretty much constant, I think.
Most video content has been at 25-30fps which is incredibly visible to most people when compared to 60fps. Those videos of cats playing with phones are rendered at 60fps.
60fps annoys me, I prefer cinematic 30fps :) High framerate us good for games, eye candy videos, but for stuff with narrative. The soap opera effect is killing me, especially when source is converted.
When you watch TV shows from the SD era that has been scanned into HD from film, you can sometimes see this. For example Seinfeld.
Occasionally things will be slightly out of focus, but it wasn’t apparent on a SD CRT so they shot the scene. On an HD screen you can see it and it’s kind of distracting.
Oh man - the same here with me - when I first got Morrowind there was an option you could enable in ATI (now AMD) cards at the time that made water look "real". I was able to enable that and was blown away - I was like "well that's it - doesn't get any more real than this". Having loaded Morrowind recently I could not believe had bad it looked lol. Makes me wonder - what are we "not" seeing right now that will make us think this way 10-20 years from now?
Hm. Many people here share their sentiments of perceiving older games (or current games) as being close to photorealistic. Personally it never felt that way to me. All games have obvious problems where they don't behave/look anywhere close to reality. If you remove the interactive element and only use somewhat massaged screenshots, then, sure certain elements are basically photorealistic and have been for a while. For example, landscapes look pretty darn realistic. Anything alive or even most man-made artefacts, not so much.
Apart from graphics most stuff in games is pretty rough. Animations are generally bad and ways before reaching the uncanny valley of "getting close"; they're still in the "abstract representation of concept" detail level. AI is dumb (largely for [perceived] player-acceptance reasons). Sound is generally poor; some games still don't use 3D sound. Physics are "abstract representation" level again, some games still have fly-through walls and vibrating items. etc. etc.
I was always so confused by people saying the latest games were realistic. They looked so horrible. Maybe it's because my point of comparison was Nintendo games where they focused more on art direction than polygon count. I read an interview with someone who worked with Shigeru Miyamoto and they talked about how he would come back with changes to specific leaves and rocks after spending time exploring a cave or forest. The attention to detail, no matter how low-poly and simply-textured, made all the difference.
This hasn't changed. New "realistic" games still range from horrible to boring-looking. I don't know if it's collective delusion or if I'm missing something.
Take this as a compliment, not a criticism when I say you've rigged your proposition by holding up Nintendo first party games, and Shigeru Miyamoto for comparison. If you were to look back at he average Nintendo game, not made by Nintendo, they are mostly unremarkable.
I should have specified first party. There were some third party games I liked, but few paid as much attention to style and detail. It's hard and expensive to get both realism and style right. I don't think it's a coincidence that most games I like for the art direction come from small indie studios who never had a shot at competing on realism.
Yeah... these comments are kinda weird. Hate to say it but maybe they should go outside more lol, computer graphics always looked pretty bad at me, even the clips in today's nvidia presentation looks really unnatural. That's not to take anything away from the technology and the massive advances it represents, just compared to reality it's still really far from fooling a human brain.
interestingly many of us feel that there's a huge gap between the past and now. But how many of us can actually verbalize what will change in the rendering of pictures in, say 3 years.
For example, I can clearly see that ray tracing produces better results. But it's a bit harder to tell how better it is, to find the words that describes how better it is. Of course, one can say that, for example, photon tracing is more physically accurate. But still, what words can we use to describe how real (or not real) a still image is ("more realistic" doesn't count :-))
Game graphics look sterile, too clean, sharp, plasticky. The real world is messier with less clear separation between objects, things blend together more, subtle shadows, surfaces are less uniform, there is more clutter, dirt, wrinkles, details.
Is that an artifact of the graphics capability? Or the art style? I think the two often get confused and many games are specifically designed for a sterile/clean look.
Not sure. It's hard to analyze as it's more of a visceral impression. It could be in part the way natural environments tend to structure themselves over time, like how we throw our stuff around in natural ways.
But also, the design often adapts to the capabilities. Games like GTA3 used to have thick fog just to hide the fact that they couldn't render stuff far away in time. You can say that's an artistic choice to give a smoggy big city atmosphere, but clearly it was a practical choice as well. Even today, game designers like to make things dark, blurry and rainy, so that the un-realism becomes less obvious.
Just watched the RT demo video and the paintbrushes for example look amazingly bad. All surfaces look flat for a better word, most noticeable it’s sterile without dust. The steam blasts are terrible as they miss all the internal swirling you get from actual vapor. They cover most of this up by moving stuff around and adding clutter which tries to keep you from really focusing on anything.
It’s basically using the same technique as hand drawn animation where as long as you realize what’s being represented you can automatically fill in missing details. However, this fails as soon as you focus on any one detail.
Honestly, it’s not bad for what it is. I mean the physics engine was probably the worst part, but as a visual demo that’s fine.
> For example, landscapes look pretty darn realistic. Anything alive or even most man-made artefacts, not so much.
Which is something really obvious with the PS5 UE 5 tech demo [0]. The environment looks great and very realistic, while the character is very stylized and looks more like a comic than realistic.
Yeah, I never understood these comments. I went from the 8-bit generation on up and never thought PSX/Saturn/n64 games looked “real”; just better and more options than before. I don’t think I ever considered a game “realistic” until the mid-late 00’s.
The biggest thing to me is that video games do not have very many independent objects, compared to reality. Go outside and look at the real world and you will see the ground has little pebbles and bits of mud and all this stuff that gets kicked around. No video game will let you inspect the leaves of plants to look for little bugs, the way reality will. Or for indoor scenes, there is no video game that accurately captures the experience of "picking up a living room with a bunch of toys strewn about".
> Go outside and look at the real world and you will see the ground has little pebbles and bits of mud and all this stuff that gets kicked around
We are getting better. Like snow deformation in RDR2, for instance(works even in a PS4).
But random bits of debris that can get kicked around - and subsequently inspected - no. That's a problem.
> No video game will let you inspect the leaves of plants to look for little bugs, the way reality will
That's an easier problem than tracking all the debris. Before you inspect, you have no idea what will be there - the computer also doesn't have to and it can be optimized away until there's an observer.
Think heisenberg uncertainty principle but for virtual worlds.
Yes, I actually thought of the snow in RDR2 as I wrote that comment! The beginning of RDR2 with the deep snow is one of the most impressive visual scenes in a video games so far, visually, to me.
You could implement little bugs without all that much trouble, true. Maybe a better way for me to think about it is more like, the real world has an incredible diversity of appearances at a small scale, that's hard to reproduce in a video game. The bug isn't the part that causes trouble, it's the fact that every leaf looks different, some leaves are oddly discolored, some leaves have bug bites on them, some got a little torn up, et cetera.
I mean, fine if you want to make an actual virtual world. But when it comes down to it, it's a video game. Why in the world would any of this be in COD or Battlefield or Fallout? It would take more time than making the actual game part of it to do this pointless garbage that nobody would care about and consume the majority of computer resources. Not to mention a lot of video games remove realistic aspects that would otherwise make it a chore to play.
There are lots of things to be improved in games, but what you're describing are improvements to a real life simulation.
Presumably you're not a native English speaker. "will" here does not refer to the future, this is a subtle grammatical thing that I can't explain well, but it refers to presently existing games.
Everything has a cost, either computational or human.
We've probably reached the point where human cost exceeds computational cost, which is to say that developing and QA testing such a feature would probably cost more money than it's worth. How many users of software would gain from such minutiae?
As someone who is most definitely not a gamer but with coworkers and friends that get big into high cost gaming systems, here’s where I see the biggest deltas between real life and the screen:
- Hair. Up until very recently hair was downright awful. Nowadays it’s acceptable-ish, but there’s still lots of room for improvement, in particular in natural motion of long hair.
- Water. I spend a big chunk of my time on the water, so I’m probably more attuned to how it moves than most. Games just don’t have it down. In particular, I think a lot could be gained by embracing it’s fractal nature: in my experience, at every human scale (mm to dam and everything in between) very similar wave patterns exist, but games tend to have just a small fixed number of “wave-layers” at various scales stacked together.
- Clouds. I can easily spend hours just observing clouds, looking at things like their shape, overall motion, internal motion, composition, edge behaviors, etc, and how they change over time. Game clouds are lacking in all these regards, particularly the time-sensitive nature of a cloud.
- Foliage. In games I’ve seen, individual plants/etc. in isolation generally look really quite decent. But the second an physical object interacts with them, they very clearly don’t respond in the right ways. There’s a lot about how branches bend and leaves rustle and more that is lost. Additionally, in groups of plants it’s often clear that some small number of models are being reused, possibly with some generated randomness added. But the variety doesn’t come close to matching what one would really see.
- Human faces and expressions. These are generally really bad, especially in normal gameplay (cut-scenes are sometimes better)
Again, this is probably all just really weird stuff I notice because I spend the vast majority of my time outdoors and only see “HiFi” games being played very infrequently. I don’t think games are worse for not implementing these, but I am very interested in what they’ll look like 10-20 years down the road.
I think these days it mostly comes down to human effort in development and not hardware limitations. You could spend hundreds of hours modelling faces or putting accurate physics on a fern but it doesn't add too much to the game and costs too much to develop.
What I think will be a huge boost is proper raytracing. You can see animated movies and blender renders look stunning but they take minutes per frame.
A good list, though I think human faces and general animation are okay when you compare with foliage, which is downright awful. A lot of the improvement in human features has to do with motion capture and 3D scanning, but getting transitions between movements is where progress needs to be made.
Foliage seems like something that will be difficult to get realistic for any close inspect, at least from a layperson's point of view.
It was one of the first prominent uses of pixel shaders in the game, coinciding with NVidia (not ATI) releasing GeForce models that supported them.
It was so noticeable because it was such a huge increase in quality compared to what passed for water in games before - usually some kind of blue-grey translucent texture. For the first time, pixel shaders produced water that was clearly an attempt to imitate water IRL, not the cartoonish representation of it.
>Makes me wonder - what are we "not" seeing right now that will make us think this way 10-20 years from now?
My vote would be cloth simulation and clothing clipping. I've yet to see a game that comes even close to doing this realistically. Imagine what happens to your sleeves when you lift your arms above your head for example, or how the plates of a suit of armor naturally slide over each other. In every game I can think of, clothing is rigidly attached to the underlying skeleton and it just stretches/clips as the character moves.
I guess fidelity of everything else has gotten much better, because I recently started noticing this and now I find it very distracting in any game that has in-engine cut scenes involving character closeups.
Maybe the water stood out since Morrowind was an ugly game even in its day. All the Elder Scrolls games have a well-earned reputation of looking like they're populated by cardboard mutants. Reminds me a bit of the old Infocom ad:
Morrowind was complicated. The characters were ugly, and animations especially horrible. But the landscapes were considered very beautiful by the standards of the day.
For me, it's fidelity of human facial animation that has a long way to come. There are teams doing a great job of it [1], but the labor/skill required to bring one high hyperreal facial rig to a game or movie seems insane. I think companies pursuing AR/VR applications will lead here.
ML creative tools stand to automate a lot of this imo.
I think Epic’s UE5 demo is really really close to reality in terms of lighting and geometry. Next 10-20 years will probably see the same technology being brought into larger and larger environments with more moving parts.
Then there’s the more obvious stuff that isn’t done well even today: skeletal animation is still lacking and feels unnatural, physics systems are still very approximate and constrained - often times most things are indestructible in games, fluid dynamics are still very slow/limited. Human models still don’t look real though, and the voice acting never quite matches the mouth movement or body language.
I do really feel like we’ve crossed the uncanny valley when it comes to natural scenery rendering. But a lot of what makes things feel real are still missing from games.
I agree, it's so close that it's actually being used as an interactive movie background with a big-ass 360-degree screen for dynamic scene and lighting: https://www.youtube.com/watch?v=Hjb-AqMD-a4
IMHO it looks fake. The lighting is way off, it quite literally looks like they're standing directly below a big diffuser - which they are. The background looks like a poster with one of these diffraction parallax effects.
It’s even harder to convince other people. Once you realize that we do not understand how to make graphics look real, the pattern appears everywhere. Yet no one will acknowledge it.
Graphics — even movie quality graphics — don’t look anything like what a camera produces. The sole exception is when you’re mixing in real video with CG elements. But try to synthesize an entire frame from scratch, then put it next to a video camera’s output, and people can tell the difference.
Also, screenshots are misleading. You have to test video, not screencaps. Video is processed differently by the brain.
10 out of 10 times, all the graphics engineers come out of the woodwork going “but actually we do know how! It’s a matter of using so and so calculations and measuring BDRF and” none of those equations work.
The reverse-problem is a pet-peeve of mine: It seems many people have been accidentally brainwashed by Hollywood into thinking that film-camera effects are signs of "realism."
So then the first-person game introduces something like lens-flares, and everyone goes: "OMG it's so realistic now", even though the truth is the exact opposite. If you were "really there" as Hero Protagonist, you wouldn't have camera-lenses for eyeballs except in a cyber-punk genre.
I think it depends on context. There are a lot of sim games where the environment is very controlled and forums are littered with people who can't tell whether the occasional picture is real or in-game. A forest is hard to render accurately but a plane in flight is pretty trivial
Forests have pretty much been solved. Look at games like Shadow of the Tomb Raider, forest looks amazing. The difficulty now is limited to finer details, like hair. Hair is still pretty much unsolved.
See? It’s almost deterministic. People really can’t accept that we don’t know how to do something.
Plop a nature video next to your forest rendering and it’ll become apparent just how unsolved trees are. And everything else, for that matter.
The precise claim is this: viewers should be able to identify a rendered video no better than random chance. If you conduct this experiment, you’ll see that real videos from actual video cameras wipe the floor.
The motion blur will probably give it away. Accurate video motion blur is computationally expensive but conceptually simple. Just render at about 100 times your target frame rate and average batches of frames together in linear colorspace. You can speed this up by rendering at a lower frame rate (e.g. 10 times your target frame rate), estimating motion, and blurring along the motion vectors before averaging the frames. You can further speed it up by using an adaptive frame rate depending on motion speed and contrast. But a lot of rendered video doesn't even try. Look at a fast-moving bright point of light and you'll easily see the difference.
(But note this is only replicating video, not reality. Truly realistic motion blur requires ultra-high displayed frame rates beyond the capabilities of current hardware.)
A lot of games fake the motion blur so badly. Racing games are often especially guilty.
If you're driving at 200 mph, and there's a car next to you also going 200 mph, it shouldn't be blurry.
Also, the length of the blur should not exceed the distance an object travels on your screen in 1 frame. In other words, if an object moves 30 pixels from one frame to the next, then the blurred image shouldn't be more than 30 pixels wide.
This feel like multiple logical fallacies. It seems funny to me to claim in advance that people are going to argue with something that was intentionally non-specific, vague, controversial, unproven, and at least partly wrong, and then when someone argues with you, claim you were proven right. It's easy to predict that someone will argue with something misleading or wrong, and that doesn't give your CG argument any credibility.
What CG videos are you considering, what specifically have you looked at? Can you show some good faith examples of the best CG forests ever made, compared to some specific nature videos? Are you talking about attempts to match a nature video, and saying it's not possible regardless of what's in the shot?
Are you looking at the best examples of CG forests lately? There are some CG full frame video examples of forests I don't believe people would reliably identify as CG, if they didn't know before hand and you left out the explosions & spaceships.
The problem is that we can't do that in 1/60 second on consumer-priced hardware, and also both scanning real objects and manually modelling are expensive.
Again, it's important to focus on video, not screenshots. Video is processed differently by the brain.
"The Dress" illustrates how easy it is to fool people with still images. Movement gives crucial context. Our visual system has evolved for millions of years specifically to exploit that context.
also, the style of photos most of us are used to seeing are so processed digitally before publication, there's a real question of what it means to have a 'real photo' to approach in the first place.
That were capable of getting pretty close isn't that surprising, because most photos have already had N layers of digital effects applied, moving them closer to renders, rather than the other way around.
Which of those images look real to you? To me, pretty much all of them look strictly like rendered images or have some aspect of them which gives away the fact that they are rendered.
With a caveat: the real images should not be cherry picked to look as close to cgi as possible but the other way around. Real and render can look indistinguishable that way too, but we want real looking cgi, not carefully arranged cgi-looking reality. The biggest giveaway is the simplicity and pristine sterility of the rendered scenes, no mess, clutter, irregularity. Just look at photos that people haphazardly snap in the real world.pick those for a blind study and cgi is nowhere near. Pick carefully lit and arranged artificial scenes with heavy post-processing, like studio shoots or hotel brochures or car ads and the difference will be less obvious.
It seems like your point is not so much that we can't technically render photorealistic scenes, but that artists don't put in the work to recreate the incredible level of detail we have in reality.
This isn't a big distinction to me? Either way, it's still true that it's nigh impossible to generate a realistic rendered every day scene. Whether that's because the environment modeling isn't there yet or the ray tracing isn't there yet doesn't make a big difference to me. Maybe the super detailed modeling will be the "last stand" prior to achieving general photorealistic renders, and the final breakthrough will be some kind of improved ML procedural generation algorithm for everyday lifelike objects.
Yes. But in principle, tech could be advanced enough that it doesn't require so much manual effort. Like, if someone takes the time an works long hours on modeling a worn book with all its wrinkles and curls and crookedness, you can probably ray trace it realistically. But you could also say that artists can digitally paint a whole scene without any 3d engine or rendering tech.
The point is, currently it only works for simple scenes and needs tons of manual work otherwise. Reality is realistic effortlessly.
In the version that is shrunk down to fit my browser, it kind of looks rendered, because you can't see the subtle details that another user mentioned. Dust, the curling of book pages, etc.
But in the full version[0], it's clearly a photo. But it has been HDR'd which can sometimes create shading that gives a somewhat rendered look.
The full version also has artifacts from the camera that ironically give it away as a real photo. At full resolution you see this characteristic noise plus noise-smoothing effect that cameras apply.
That effect would be much easier to fake with a digital filter than rendering the entire complicated scene from scratch, though. Adding noise and noise-smoothing is simple. So if that's your heuristic, then you're gonna be fooled by digital renders that just incorporate that simple to implement effect.
Yeah, absolutely. It also shows though how you can easily accidentally bias a blind test. If an artist did manage to create a perfectly realistic render but forgot to add in one of the subtle unrealistic effects that cameras cause, people would be able to tell which images were photos.
(Although, maybe you could argue that if the challenge is to make photorealistic renders rather than realistic renders, nailing the camera artifacts is part of the challenge.)
Yeah, the goal has to be photorealistic renders, because what would the control data for the test be if not photos? You can't test one image on a screen vs "what real life looks like". The one on the screen is obviously the rendered one then every time.
It's not real multiple exposure HDR, it's just whatever default faux single exposure digital HDR the Pixel phone does. If it were real HDR the photo wouldn't be so blown out and over-exposed from the sunlight on the left side, particularly on the leaf.
I think Instagram is actually helping this from the other direction. My brain processes the two Jeep pictures as real but with a heavy-handed Instagram filter applied.
Do any of those actually look real to you? I'm not saying that a convincingly realistic render is impossible, but those are all obviously synthetic to me. Maybe that's just a lifetime of photography at play.
Meh...all of those are obviously renders. There's not enough contrast in the lighting in most of the interior shots, and overall surfaces just look too smooth.
Your comment seems to be ignoring the fact that today’s graphics look better than yesterday’s, i.e. the main point of the comment you replied to. The monotonic trend toward realism has continued ever since computer graphics was invented 40ish / 50ish years ago. VFX and games from 10 and 20 years ago look much less realistic that today. In addition to stills, there is an increasing number of videos and film scenes that are getting harder to distinguish from shot on film. Are you saying the trend will stop? If so, why?
> we do not understand how to make graphics look real
> none of those equations work
Why do you claim this, and what do you mean exactly? Surely you aren’t claiming we can’t make any graphics look real at all. Some people are getting pretty good at understanding how to make some graphics look real, even if we aren’t. The math for approximating materials and lighting is getting pretty good these days, and the equations have been changing over time. It’ll continue to improve, but it’s certainly not the main thing holding CG back today. IMO less realistic CG comes much more from low resolution or low fidelity sampling / geometry / textures / animation / input data in general. Most of that is due to lack of time & money.
I’d say we understand how to make graphics look real in many cases, but not in all cases. We understand how to make good looking graphics, but how to do it quickly and efficiently enough to be practical and done with limited time and effort is an open problem.
So I've dabbled a bit in graphics, and it feels like a content problem. Sure, you can have an artist pump out some models and textures. But for every level of detail increase, the artist must spend ~3x the time of the previous level. So for example, even making just a simple a dirt road look _really_ good and photo-realistic would involve much more time (a week? a month?) than one can reasonably spend on a commercial project (where you have bajillions of assets to worry about).
I didn't study this movie frame by frame, but I feel the new lion king movie was photo realistic. It seems like that given enough time, human and computer, you can achieve a great result.
When people are saying something is photo-realistic, they aren’t comparing to reality, they’re comparing to photos or videos as viewed on the same display.
By some metrics, even extremely expensive modern hardware is very far from reality. Pretty much all games show Sun occasionally, to reproduce same luminosity at 1m distance you need kilowatts of light (assuming 180 degree viewing angle; the surface of 1m half sphere is about 2 m^2 and the flux of visible spectrum is about 550w/m^2), these levels are simply unsafe for home use. For example, such display is going to ignite or at least melt a keyboard if it’s too close.
Similar for dynamic range. Reality has extreme dynamic ranges, i.e. both very light and very dark areas on the same scene.
At least modern hardware generally delivers realistic resolutions, and color accuracy.
> At least modern hardware generally delivers realistic resolutions, and color accuracy.
I don't know, look at some flowers... most screens can't show their colors. Stuff like the intense, super-saturated reds and purples are basically impossible to get right in sRGB, with very obvious artefacts, yet in real life there are no artefacts, there is texture, detail and color, where the picture only has a smear of four different reds. P-3 and Rec.2020 might reduce the issues there, as would 10 bit color.
I remember the progress from CGA to EGA; there was that odd bridge of Tandy 16 color; and then, boom, VGA, which looked so magical compared to what came before. Even when I did gray scale VGA on my IBM PS/2 50z it all just felt like a big jump had been finally taken.
Then down the road came 3dfx with Voodoo and that to me was the next great leap forward. Each iteration has been leading to ray tracing which is the next great leap.
Now just for screen tech to become as affordable as the cards that can drive them, the LG OLED we have is stunning but that is "just" 4K.
Weeeeell...the first 'great leap forward' was the Verite V1000 and vQuake a couple of months before the Voodoo 1. Real 3D looking Quake with good frame rates! A bunch of us jumped on that bandwagon...and prolly shouldn't have. The V1000 was great for vQuake, much less so for anything else. By the next Xmas, we all had Voodoos.
> I considered certain game elements like fire and water indistinguishable from reality. Now it looks like pixelated garbage and I ask myself how I could have ever thought that.
Imagination. Same as the people who grew up with the original Atari and Sinclair Spectrum and Commodore 64.
That’s interesting. Even if I look at current state of the art graphics today I don’t think for a second that it’s anything close to indistinguishable from reality (not to say it’s not good or impressive).
I guess he means how reality looked displayed on the TV screen. I think we did go backwards in that regard, at least for a while because resolution went up in some sudden jumps leaving the level of detail in the scenes lagging behind. But we're almost there now. I recently saw a dashcam video of a racecar on the news and wondered why they were showing a video game. It took ages for me to realize it was actually real but with some HUD overlays.
I remember when I first saw a DVD on a good screen. It was the first Matrix movie in a friends basement. I was so blown away by the quality jump from VHS.
The nice thing about 3D graphics is you can play those old PS games on an emulator on a modern machine at a higher resolution. There's a law of diminishing returns of course, and the older textures will look blocky even after being upscaled, but it still looks better than the original hardware.
I remember playing Final Fantasy 9 entirely on an emulator and how much better it looked compared to a "Real" PlayStation.
As always, wait for the benchmarks before deciding to buy (or return). My guess is the performance improvements are the biggest for raytracing, which I personally don't care for. And let's not forget the huge power draw, requiring a new power connector and a 3 slot cooler.
8N manufacturing process is presumably Samsung, which will probably be beat by TSMC 7nm.
> As always, wait for the benchmarks before deciding to buy (or return).
Why would you keep a 1200$ RTX 2080 Ti or a 2500$ Titan when you can get at least the same perf for 50-75% of the price with the new products, and much better RTX perf ?
This assumes that with RTX off the new gen won't be slower than the old one, but I think that's a fair assumption, and if it isn't, you can always return the 30xx card in 2 weeks, and buy an used 2080 Ti or Titan on ebay for pennies, once the market overflows from people upgrading for the 30xx series.
People were still asking for 900$ for a used 2080 Ti on ebay this morning, and the 3080 700$ price just destroys that offer. Many owners are going to try and dump these cards for as much as they can get in the next two weeks. I wouldn't buy a used 2080 Ti for more than 250$ today. In one month, these cards are going to sell used for 100-200$. If AMD can only match the 2080 Ti perf, they are going to have a very hard time pricing against the used 2080 Ti market.
> And let's not forget the huge power draw, requiring a new power connector and a 3 slot cooler.
That's only for the 3090 IIUC. All other cards were announced as being actually much smaller than the 20xx series ones.
True, the new generation offers better price/performance vs the previous generation.
Do note that the 3080 is still 2 weeks away (17 sept), 3090 24 sept, 3070 in October.
I would consider getting a non-founders edition though, in the past other brands have had better/more silent cooling for pretty much identical pricing.
Edit:
While the 3090 (350W) is the only one requiring a 3 slot cooler, all 3 use the new 12 pin power connector. The 3080 founder edition power draw is still 320W, vs 220W for the 2080.
Pretty bold claim that 2080TIs are going to sell for $250 in a month, lol.
Certainly the market for them will take a beating compared to new prices now. But I can't imagine it collapsing like that purely based on the supply. How many people are really going to drop their 2080 just because a new thing is out there?
I'd love to be wrong, as an original 2070 owner ;)
>Why would you keep a 1200$ RTX 2080 Ti or a 2500$ Titan when you can get at least the same perf for 50-75% of the price with the new products, and much better RTX perf
Well for one, keeping a card you already bought is free, buying a new one costs money.
Why are you all getting so excited about the prices given in a paper launch? Wait until you can actually buy one at that price, which I'm guessing isn't going to be until well into 2021.
I'm not sure if I'm missing some context but I'm assuming you are outside the return period. But anyway, I'd be really beyond surprised if these cards are actually available at these prices anytime soon.
Also if find a 2080ti for 200 before the year 2022 let me know I'll buy you a beer.
Hardware Unboxed tested an RTX 3080 and it seems like the performance improvements are not just in ray tracing: https://youtu.be/cWD01yUQdVA
It looks pretty impressive.
Regarding the new power connector: All the partner cards showcased by Gamers Nexus have the regular two or three 8-pin connectors and the FE cards seem to include an adapter. Nvidia states that a 750W power supply is required, so depending on the CPU even a 650W should be fine.
> My guess is the performance improvements are the biggest for raytracing, which I personally don't care for.
I still don't understand the raytracing craze when it comes to games. Years ago when I wrote my first raytracer I might have got excited about it. But we have advanced so much with rasterization that I don't really understand why this is something we need.
Raytracing complexity is tricky and it is likely to prove challenging to do in a game with consistent frame rates. Soft shadows are expensive even for hardware.
I would be more excited about some other global illumination improvements, like photon mapping.
RTX cards don't render everything with ray tracing. They use it for light calculations, for example it can be used exactly for global illumination.
> why this is something we need.
The answer is simple. Because there is a limit to how realistic we can make the lighting and shadowing using the old models. We are almost at that limit. Using ray tracing removes that limit.
> challenging to do in a game with consistent frame rates
Try Quake II RTX. It is very consistent right off the bat, and if you turn on the resolution scaling, you can get completely consistent framerates (within 2-3 fps continuously). While it might not be a game with the most advanced graphics, it does have variety of scenes, and there is no issue with the framerates being "inconsistent".
Many newer games are designed to show off raytracing effects. Cyberpunk 2077 has lots of neon lights, shiny things, rain puddles, etc for light to bounce around, and raytracing makes it look a lot nicer than when the shadows are approximated in traditional ways. Minecraft ... ye gods. The videos I've seen look amazing. The other nice thing these cards have is the DLSS 2.0 tech, which is supposed to let it render at higher resolutions by _making things up_ (?) based on ... something buzzword-laden. The effect, though, is that you get similar or better quality at much higher framerates, because lossy techniques are still nice to look at.
"Need" is a strong word, but I'm sure someone will make a game where the shadows _matter_ for telling the story, or solving puzzles, or even just making the game world look more realistic and believable.
DLSS renders the game at a lower resolution then upscales it to a higher resolution using AI to modify the final image. The models are trained from super-resolution renders and used to be trained per-game but are now generalized. They're designed to upscale textures, denoise, fix aliasing and overall increase quality based on that high-end source data so you get the same output without having to actually run at those settings.
> "Need" is a strong word, but I'm sure someone will make a game where the shadows _matter_ for telling the story, or solving puzzles, or even just making the game world look more realistic and believable.
Yeah there is no way they sell a card that does 200% the fps of a 2080ti in "normal" (non raytracing, ultra quality) averaged over multiple titles. If a 2080ti does 100fps without raytracing and 30fps with, this might do 110fps without raytracing and 60fps with, for they'll a "100%" increase but most would consider that 10%.
I was wondering if all benchmarks use the OptiX features or only the CUDA features. In software like Blender this makes a huge difference. OptiX can be twice as fast as CUDA.
So of benchmarks don't use the OptiX capabilities then they will miss a lot of available speed.
The product page linked elsewhere in the thread[1] has photos of the 3070 with the 12-pin connector and mentions that all Founder Edition cards include an adapter:
> To that end, our engineers designed a much smaller PCB, shrank the NVLink and power connectors, and still managed to pack in 18-phases for improved power delivery. Don’t worry, we included an adapter that allows Founders Edition cards to work with users’ existing power supplies.
Well, 2 generations ago the MSRP of the 1070 was $379 and the 1080 was $599, so I'd say we've just gotten used to the higher prices which they successfully normalized.
That's sorta fair - we're assuming that these cards won't be marked up like crazy either - I imagine the MSRP will be wrong here too. I bought a Gigabyte 1070 for ~$450 USD at launch.
Agreed, seems like Nvidia is preemptively trying to fend off AMD with an extreme show of force. This level of performance/pricing is much higher than even AMD's most optimistic estimates for performance/efficiency gains for their next generation.
I couldn't find pricing for 3090. I seriously doubt it's less than the current version of Titan. It'd also make sense - lots of DL research is currently done on 2080Ti. Take that away and people will buy a more expensive SKU just to get more VRAM.
If this pricing holds up though, I need to get a large NVDA position, because they'll sell at TON of 3090. I'll buy 8.
Yeah, naming is confusing, the RTX 3090 is the non-HPC "Titan" model of this gen, and, if I understood correctly, is 1.5x faster than the previous gen RTX Titan, which costed this morning 2500$.
There is also a previous gen HPC "Titan V" with HBM memory, but AFAICT no Ampere Titan card with HBM2 was announced.
There is a rumor about 48GB version of RTX3090 replacing the original Titan RTX with a similar price. That would make sense for Deep Learning workloads as 24GB is already too small for attention-based models.
It's 50% in ray-tracing performance, not raw performance, and it doesn't have big memory bandwidth improvements. It's not the Titan successor, it's the 2080Ti.
You can see this because the chip name is GA102, the 2 at the end indicates that this is a cut-down chip.
All previous generation Titan models have been cut down chips as well TU102, GV102, etc.
With one exception: the Titan V card that I mentioned which comes with the GV100 chip. But as I mentioned, that one targets a very different market segment with HBM memory, which the "RTX 3090" obviously doesn't target, since otherwise it would come with HBM2 memory like all the GA100 products.
Extremely confusingly the Ti stands/stood for titan as well, while at the same time Nvidia also released cards with the name "Titan", like the Titan X, Titan Xp, Titan V, Titan RTX etc. When people say "the titans" they may refer to either the "something Ti" cards or the "Titan something" cards.
The video[1] mentions that it (3090) was made for users in the Titan product segment, and it's introduced as a Titan replacement. The Ti models will probably come later as usual.
All this on Samsung 8nm (~61 MT/mm2). They didn't even feel the need to use TSMC 7nm (~100 MT/mm2). Probably keep the price down and to reserve capacity at TSMC for the A100.
This is like the anime hero/villain (depending on your perspective) equivalent of fighting at half power.
I don't have a source on hand, but my understanding is that Samsung significantly undercut TSMC on pricing, likely to stay alive in the high end fab business. Samsung is probably making extremely low margins on these, hence the lower than expected MSRP.
On the slide in the presentation it did say something like Samsung Nvidia Custom 8NM process, so perhaps Nvidia made such significant contributions to the process that it's not really Samsungs process anymore?
TSMC's "customized" 12FFN process just had a larger reticle, so the burden should be on Nvidia to explain any customization. I don't think they deserve any benefit of the doubt here.
Yeah, impressive for the node. TSMC's latest N5 process is 173 MT/mm2, which Apple is using now and AMD (probably in N5P form) will start using next year. EUV is really a big step up.
The 20xx series was very obviously the "tech development" release and was a terrible value. It was first gen ray tracing and thus actually unequipped to fulfill its promises. The 30xx series looks to be much better and is probably finally worth the upgrade from 9xx and 10xx equipment.
It was the 20xx series that had insane pricing (in a bad way). nVidia had no competition and priced things at the highest markup the market could take, segmenting it by ML, cloud, mining, gaming etc.
Not sure what their competition now in each vertical, but apparently they believe that they need a lower price point.
I'd wait for in-game benchmarks on high end monitors before I believe that comparison from NVIDIA. Even if the specs are better games are still optimized for older gen cards.
Yes, but we've all been burned by GPU marketing in the past. Notice that they're only remarking on RTX performance. It's entirely possible that this card has a lower geekbench score or FLOPS output.
I'm still sitting on a GTX 970 and was waiting for the RTX 30 release so I could buy a used RTX 20 at under $500. This pricing means that it's worth it for me to just jump straight to the 30 series.
I recently got a 2070 super OC. It really struggles with, say, COD:MW at near 4k Res (3440x1440). Indeed the campaign with RT on doesn't get more than ~50fps. Kinda disappointed really. Get about 100fps in multiplayer with everything turned off/low.
But the cost of a 2080ti was ridiculous especially considering the open secret of its impending obsolescence.
AMD really need to up their marketing budget. From the zeitgeist I've no idea where their lineup sits comparatively. No wonder they only have 20% market share
AMD still don't offer a compelling product for the high-end, which Nvidia is taking full advantage of. AMD cards are the most economical on the market and have their own advantages but they're niche (Linux support is much better for example).
AMD got on top of Intel by creating hardware that delivers. Just a few years ago AMD CPUs were economical, but the performance was abysmal, especially in single-threaded applications by comparison. They didn't make a product for the high-end.
I am eagerly awaiting what the next series of AMD cards are going to be able to do. They're talking a big game for sure. But Nvidia has a big software advantage as well as a hardware advantage on AMD and that's likely to be a sticking point for me personally on my next purchase. Nvidia spends a lot of resources on working closely with developers and providing them support they need to take better advantage of the hardware with nvidia-specific features. AMD doesn't seem to do the same, and has had much higher profile issues with their drivers in my experience.
All that said, I hope AMD can provide a product to truly compete at the high-end with Nvidia, to hopefully drive prices down as GPU prices have gone up dramatically on the high end.
About 4 months ago I picked up a 2060 Super for my ancient gaming desktop while I am waiting for the 3080 or 3090 to finally build a new gaming PC. Ideally I'd have a 4th gen Ryzen equivalent of the 3900X.
The main reason my 2012 build of a PC is holding up ok was PCIE 3.0 support, so for me PCIE 4.0 is a must. The only thing I ever upgraded was the GPU and HDD. Went from 670 -> 980 -> 2060 Super. The i7-3370k (OC to about 4.3Ghz) has held up ok. The 16GB DD3-1866Mhz is slow however. Switching from a 2TB 5200 RPM drive to a 2TB NVMe SSD (Samsung EVO 860) for which I had to get a riser card since the ASROCK z77 Extreme 4 doesn't have m.2 made a huge difference also. When I upgraded to the RTX 2060 Super I ran out of PCI 3.0 lanes, and as a result my SSD is running slower, but that's ok.
Suprisingly, I can play the new Microsoft Flight simulator just fine in 3440x1440 resolution on High Settings. Assassin's Creed Odyssey runs well in 1440p Ultra at ~60 FPS.
I just went and checked my RX 5700 receipt and I paid $212 after tax last November. After selling the included Borderlands 3 license because I'd already bought it, net price under $200. GPU value of the century right there.
But as a lighting nerd, I do want the raytracing...
Ah, you're right. It was a good sale but not quite that crazy - still under $300 total. Combo deal from Newegg and for some reason they invoiced the two parts separately, so when I searched by "5700" it only found the first one. Went back by order number and found both invoices.
Still, less than $300 for this card has been a great bargain. I've since picked up a 1440p screen and it's happily chugging along. Been most of a year and they're still selling for substantially more than this.
Yeah, the price inflation on midrange cards was pretty crazy. My GTX 970 was around $375, and the 1000/2000 cards were way more expensive. The RX 5700 was the first worthwhile upgrade in a similar price range. Lack of competition paired with cryptocurrency miners.
I see a bunch of both cards and they're both around 400 bucks, give or take... but I kinda figured putting a link to a view into a selection of newegg inventory would end confusing at best :facepalm:
There's been significant stocking issues on PC components since the pandemic began and thus prices have gone all sorts of crazy. You absolutely shouldn't spend that much on an RX 5700; the performance isn't remotely there to justify that price.
$290 + tax, and $30 back for reselling Borderlands, so effectively $260. BL3 was only a few months old but I think the value was tanked a bit by being an Epic Store key, it was free so I can't complain.
The launch prices were $350 for the 5700 and $400 for the 5700 XT (after the pre-release price drop response to Nvidia), so $260 was $90 under list price, or 25% off.
And really, even finding a GPU for list price instead of being marked up for mining ethereum (or whatever else) was a bit of a miracle compared to the last few years of GPU price insanity.
I bought a 2060 two weeks ago for $280. I don’t want to drop $499 on a 3070, very psyched about the 3060. Just seems like that won’t be until late this year.
>Leveraging the advanced architecture of our new GeForce RTX 30 Series graphics cards, we’ve created NVIDIA RTX IO, a suite of technologies that enable rapid GPU-based loading and game asset decompression
It can't both bypass the CPU and have decompression unless it is decompression on the GPU. I'm not sure it is dedicated decompression hardware, or if it is using the normal GPU compute.
> Specifically, NVIDIA RTX IO brings GPU-based lossless decompression, allowing reads through DirectStorage to remain compressed while being delivered to the GPU for decompression.
Still not sure if it'll use fixed-function decompression units on the GPU or if it's just compute shaders, but it's decompressing on the GPU.
probably because of, oh I don't know, the text right next to the 3070 dot saying "faster than 2080 Ti"...
(and I'm really only speculating here)
It probably won't be like, night and day faster than a 2080 Ti of course, it's going to be the same bracket judging by the chart, but I'd expect it to usually edge out the 2080 Ti by a couple percent based on the text there.
People are speculating that the 2x numbers come from RTX and/or DLSS, but they also doubled the number of CUDA cores so its possible its going to be an actual raw performance gain.
I think its wise to be skeptical until independant benchmarks are available but I would be surprised if this didn't end up being the biggest performance increase in Nvidia's history for a single generation, just like they say it is.
All of the performance claims are extremely hand-wavey.
They are using the deep learning and compute features to equivocate about resolutions and framerates. 8K@60 isn't real, but with DLSS activated it is.
I don't know about last gen, but that was the same as the generation before it - the 1070 beat the 980ti. It's the cycle they do, and as a gamer why I'll never buy more than the x70 card.
I'm going to assume that's because it has twice the raytracing power so it's with all the eyecandy turned up (Until I see something that says otherwise).
> Everything was awesome except the 3090. $1500?!?!?!?! That's a big jump from the 2080Ti pricing of $1200.
Since its a Titan model (for machine learning work, not for gamers), and the last gen "RTX Titan" costs 2500$ today, its actually a big jump, but in the opposite direction. Almost half the price...
Seems like the expectation on Reddit is that the 3090 is more of a Titan replacement rather than a 2080Ti replacement and that a Ti variant will come later with 20gb of mem.
It's not an expectation from Reddit, this was said explicitly in the presentation - that they used to make the Titan for people who wanted the best of the best, so here's 3090 for that market.
From what I heard, nvidia wants to simplify naming, having Ti and Super along side the numbers makes it confusing for the consumer (in their opinion), so they want to go away from it. I wouldn't be surprised that the rumor is true and that they will ditch the Ti/Super suffixes.
That is fair. It is a bigger jump from the Ti. Though most of the rumor sites were suggesting $1400. The 20Gb 2080 variant seems to be a really recent rumor though.
What I did find interesting is that it does seem like the $1499 on the slide could have been mistakenly shown. They didn't verbally announce it and other than that one second avoided talking about the price.
3080 ti comes later. They said at the presentation the 3090 is the new Titan so its actually a massive price drop.
The 3080 TI will probably be out next year if they follow typical patterns and it will have slightly better gaming performance than the 3090 at a much lower price.
GPUs are still overpriced because of Nvidia's monopoly on high-end cards. The fact that we're so normalized to this after the 2xxx series is kind of sad. It's like praising Apple for a phone that's "only" $999.
How are you defining value? Sales for the 20XX were "disappointing" and so obviously the market considered these cards to be overpriced. However, the reaction to the 30XX announcement seems to be extremely positive.
Isn't the "value" of a GPU simply what the market is willing to pay for it? I bought a 1080 in May of 2016 for $599. According to the BLS CPI inflation calculator, that's approximately $647 in 2020. $499 for a card that is double the performance of a 1080 seems like a great value to me, personally.
The rumour mill has the top end big navi at 2080 ti levels of perf. If those rumours and nvidia's claims here are accurate, they're out of the high end segment. again.
I wonder what the price would've been without AMD dipping their toes into the GPU market again.
I very much assume Nvidia slices price, to not only compete but also crush the AMD GPU market. Nvidia's innovation is crazy huge, but they probably don't want to be in the same situation as Intel where people regard the big dog as slow, profit-extracting. Nvidia will do it's best to eat as much market share with better products over AMD('s GPU.)
Based on performance claims from AMD, as well as PS5 and Xbox, I would expect to see an AMD card which matches the 3070 in rasterization for around the same price, maybe a little less, but with less of the fancy software features. If you look past the marketing, you can see that Nvidia's improvement from a performance/watt perspective has only increased by ~20-30%.
The 2080 TI are going to overflow the used market, at 100-200$ at best, so... if AMD can't top that significantly, a used 2080 TI might be much better value than anything AMD can offer.
That's overly aggressive. If a 3070 is really the same performance as a 2080 TI then the 2080 TI will likely go used for $400.
People don't see graphics cards like they're brake pads. They don't wear down the same way. $150 (no tax) is still $150 less for what is essentially the same thing.
many things is: ray tracing, tensor cores (for DLSS), and normal shaders (which is essentially everything).
Even if you don't value ray tracing, DLSS is the difference between playing at 1080p ultra settings, and doing the same thing at 4k.
Now, if you say you don't value Ray Tracing and DLSS, then I'd agree that it makes little sense to upgrade from a 2080 Ti to a 3070 or up. But DLSS is quite useful for any 3D application (CAD, 3D modelling, gaming, ...), and the tensor cores it uses are useful for ML, so if someone is not doing either of these with a GFX, I wonder what do they use the GFX for where DLSS makes no difference. Maybe as a heating stove :D
> Now, if you say you don't value Ray Tracing and DLSS, then I'd agree that it makes little sense to upgrade from a 2080 Ti to a 3070 or up.
I think the thing is though - how many people use DLSS and how many people use ray tracing? I'm going to guess very few. This is all the games that support ray tracing and/or DLSS. https://www.rockpapershotgun.com/2020/09/01/confirmed-ray-tr... That list is horribly short. I don't even play any of the games on that list!
Again, the amount of people using the tensor cores and what not is trivially small. I'll be excited if the benchmarks come out and show a 50%+ gain across the board with all games. Until then - I remain skeptical. This just seems like a price cut on a ridiculously overpriced amount of cards. The fact that the 2080 Ti is still $1200 after 2 years of being out is atrocious. Nvidia went ultra-capitalist with their last generation.
It's targeted for a November 2020 unveil[0] so we'll have to wait to see what they're coming out with, but I doubt it'll stack up to the 30xx series on both price and performance (at least, based on these marketing slides).
It has to be out before the new consoles, which are reportedly coming the first two weeks of November. I think the consoles are being kept secretive on AMD’s request as well.
I have some CUDA code that I need to run as fast as possible, so if I was going to blow $1500, my use case would imply that I should go with two 3080s.
However, I also play around with deep learning stuff, expect to do so more in the future, but don't currently follow it so closely.
Would someone care to ponder on what difference they think a 24GB gpu vs a 10GB gpu will have as a tool for deep learning dev over the next 3 years?
For what it's worth, I'm a computer vision guy, but I did have a play with DeepSpeech earlier this year.
Two 3080s = $1500 + number of hours times electricity cost (for me, this would be about $0.14/hour)
A K80 from AWS (close enough to the same RAM) = $0.45/hour (spot pricing) or $0.9 (on demand)
I've been able to use spot instances for basically all of my AWS hobby work, so I'll use that.
The crossover point where it's worth buying is where 1500 + 0.14h = 0.45h, about 5000 hours or 30 weeks of training.
For hobby work, do you expect to have it training for 20% of the next 3 years? That's all without considering the fact that you might want to upgrade to a newer card, AWS's prices will likely fall, and you might want a card with more RAM sometimes, but not others, or even a multi-GPU setup sometimes.
I used to spend money on fancy cards and machines and justify it with my hobby learning. Now I just let somebody else do the heavy lifting and pay them rent. It's spoiled me. You go from spending a few nights trying to get a thing performant to throwing a burrito worth of money at amazon to just parallelize the dumb thing on a massive machine and having it done by bed time.
I also used to justify buying the fancy card with my gaming hobby, but then when I actually did have a long running training job, it'd be super frustrating because now I can't use the desktop for anything else for a few days until training finishes.
I have to admit you're right. My case is a little different, where my jobs are about 1h long and it would take me longer to copy all the data to AWS than running it locally.
If you transition your workflow to AWS, and do preprocessing and storage and training in AWS, it avoids the issue. As a benefit, storage in s3 is dirt cheap compared something like dropbox, and archival is easy and cheap enough to just forget about. I feel weird like I'm evangelizing amazon, but it's a workflow i'm very happy with and trying to streamline to take even more advantage of.
Well, they can be 40 GB fp32 models technically, but translating a model trained in fp32 to fp16 is not trivial (trust me, we’re working on this right now for a model). But remember that training the model requires a lot more memory than just the model parameters, because you need to store the gradients as well.
What the other commenter said is absolutely true, but what's even more important is that 3080 doesn't seem to support SLI so you're "stuck" with one 3090 with that budget (which seems to be the only card supporting SLI this gen).
Not an expert in ML, but I don’t think CUDA uses SLI at all.
SLI is specific to rendering. Depending on the workload, it’s sometimes makes perfect sense to split GPGPU jobs into multiple GPUs. An extreme example of that approach is crypto-currency miners who sometimes use a dozen of GPUs in a single computer.
The only limitation, the working set used by each GPU needs to fit in VRAM of that GPU, otherwise GPUs gonna bottleneck on I/O as opposed to compute, will be very slow. For ML, this means the setup of two 3080 GPUs will be limited to 10GB model sizes.
It's not impossible to distribute training across multiple GPUs, but it's certainly not straightforward. And if you want activations for an entire 4K image in the same model, having a lot of memory on one card is your friend.
I'm fairly lay but imo GPT-3 demonstrates pretty soundly that huge models are no magic bullet- it's got >2x as many parameters as the human brain has neurons and it can't do long division. Dogs and other animals get by just fine having less than 1% as many neurons as humans.
Even a billion parameters is a huge model, and a factor of 2.4x increase is not going to make a tremendous difference in your performance. In particular the data-heavy nature of vision stuff means that you'll be bottlenecked by training more than memory, AFAIK (again, lay).
This is a misleading comparison. You are comparing a massive model with huge models. What you should be comparing are big models vs medium models that a single consumer GPU will fit. And - you don't need to take my word for it, there's tons of papers - the bigger models definitely perform better.
It can't do long division because it literally can't read the number inputs, due to the way text is encoded (BPE). That it manages to learn any sort of arithmetic despite that is pretty impressive to me.
This might be an unpopular opinion, but is the new RTX 3070 at $499 and RTX 3080 at $699 really such good value?
Sure it's being marketed as 70-90% better performance at the same price but the value proposition is mainly in comparison to the Turing generation (20XX series) versus the Pascal generation (10XX series). For example of 3070, Nvidia moved the price anchor for the XX70 series from $379 to $499 so consumers are essentially paying for a XX80 card - and should expect XX80 performance. Nonetheless still impressive for the 70-90% performance gain, just perhaps not as much when price-adjusted.
On a separate note, I'm curious about how much of the performance gains is attributable to the node shrink (12nm to 8nm) vs. micro architecture vs. software optimization (e.g. DLSS 2.0 vs. 1.0).
If you can get something better for the same money or a little more, then sure they are good value.
Even if you want something not as powerful or new, these will just send all the older card values down, so for that - these are great value.
Now if you was a TITAN buying type, the GTX3090 is definitely good value.
As for node shrink and how much is it a factor; Would need to factor in clock speeds and memory bandwidth, then need to compare exact comparable features. It's a muddy path. More so with the bigger gotcha that the 12nm node was TMSC and this 8nm node is Samsung, so hard to say, but over the months, reviews will have a good stab at it. Though for me, GamersNexus on YT would be the review I'd be looking for in the weeks ahead.
8K gaming is the ultimate gimmick. Even in the ideal conditions they set up for that demo, I'm doubtful those gamers have the ability to detect a significant improvement over 4K.
That's fine. Its really hard to get good frame rate in 4k today. A lot of that isn't just the GPU (eg: poorly optimized games, CPU, I/O, etc), but if it can do 8k 60fps reasonably consistantly on paper, then it can do 4k@60fps+ consistently for real (assuming nothing else is the bottleneck).
That makes it worthwhile. That's personally what I was waiting for before upgrading my 1080 GTX TI and my monitor.
I was surprised and disappointed to see no mention of VR. Seems like a key area where significantly increased GPU performance can make a big difference, but it doesn't seem to be in sharp focus for nVidia. It may accentuate the direction Facebook is already taking to move more and more to mobile chipsets but that should be something nVidia would want to minimise if they were really smart about it.
We are at the stage where nothing is enough for either fps or resolution. I use a vive original and its painfully blurry and thats 1080p per eye and a very low fov. For a high fov lens we will probably need 16k per eye at 144hz before people will consider it good enough like 4k for monitors is now. We don't even have the cables to push that much data currently, let alone render it.
I expect to reach this we will probably have to start using foverted rendering.
There already exist 1/2 8K VR headsets though (full 16:9 4K per eye), and you can keep seeing a difference there up through around 16K per eye and higher refresh rates than they showed.
In my experience, high resolution is only relevant to games where the camera isn't constantly moving.
Think DOTA 2 or Civilization 5. They both look amazing at 4K and I bet would look noticeably better at 8K.
Especially in those two, the games assets have enough detail to allow zooming in all the way from a bird's eye view to a first person view. As you crank up the rendering resolution, there's plenty of "real" detail available in the models and textures. You could push these games to 16K and still get more quality out of them.
Agreed. and in high movement games (mostly first person shooters), you want higher framerate over higher resolution any day. 120+ fps is a night and day advantage over 60fps.
8k on a normal sized monitor is a gimmick, but 8k gaming on a 98"+ tv is not. These TVs are mega expensive right now, but they might not be in 2 years. Nvidia is ahead of the curve.
Agree that 8K is a bit overkill, but it'll be nice for 60FPS+ 5k since 5k doubles 2560x1440 exactly and makes for a better multipurpose use resolution than 4k does at 27".
Doesn't matter much in games, but it makes a difference on the desktop. Don't know about Windows but under macOS and Linux integer scaling generally works more cleanly and predictably than fractional scaling does.
Just compare how much Nvidia is pushing the limits as the industry of leader of GPU’s, to how Intel seems to be playing catch-up reluctantly as the industry leader of CPU’s. Leadership matters.
Nvidia isn't a fab. AMD isn't a fab. TSMC is the fab that beat Intel. TSMC's customers, like AMD and NVidia, are beneficiaries.
So far. As customers of TSMC, in the long term it behooves them for TSMC to have competition.
Further, "leadership matters" is a somewhat ironic complaint given that Intel ran face-first into a brick wall precisely because they were leading. TSMC placed conservative bets on the next node and Intel placed risky bets because they needed the extra risk/reward to maintain leadership. Intel's bets failed (in particular, cobalt wires and COAG). They chose "leadership or bust" and went "bust," at least for now.
The whole reason AMD are able to crank out 128 core CPUs is the CCX architecture - the one people laughed at. No TSMC there. Not to mention other innovations like Infinity Fabric.
In ampere for instance, there are so many innovations, like PAM signalling, 2x shader instructions per clock, DLSS, RTX Voice.
TSMC beat Intel, sure, but that is not the main reason for why Nvidia and even AMD are leading the industry. In fact, ampere is on Samsung 8n.
'gallerdude didn't mention anything about fabs. Also, these GPUs are fabricated at Samsung, so your comments about TSMC are mostly irrelevant.
If the fab was all that mattered, AMD GPUs would be dominating Nvidia, since they have been shipping GPUs using TSMC 7nm (a superior process to Samsung 8nm) for over a year.
as someone who put failing green IGPs on notebook pcbs into the oven in 2013 for a reball reanimation I feel some irony when the chef himself is at the old trick.
backsplash! That's a word I did not know. Anyway, I couldn't help to notice the backsplash too and the kitchen enclosing, it looks optimized to collect grease and make it hard to clean :-/
If their performance claims are accurate, AMD has a huge hurdle ahead of it, as the rumour mill only had them drawing even with the 2080 ti with big navi.
I hope AMD can pull it off, as I am really hoping to make my first red box build. That being said, the performance:cost ratio of 30xx is mindbogglingly attractive (assuming the reviews back up NVIDIA's claims).
Digital Foundry's initial analysis mostly bears this claim out.
They found FPS increases of between 160-190% for a bunch of recent games featuring both RTX/traditional rendering, and about a fixed 190% in Quake RTX (which is exclusively RTX rendering).
Those titles were all cherry-picked by Nvidia though. I would be very hesitant about the data used. A lot of DLSS used too - so, not even true 4K.
On top of that - if you only care about 4K gaming then fine - but what if you're more of a 1080-1440p with higher refresh rate person?
Personally, I have a 1440p display that goes to 144hz. I'd much rather have 1440p @ 100+ FPS than 4K at 60 - but we don't know what these new GPUs will do.
The rumours I've heard have them beating the 2080Ti, but not enough to be competitive with the top Nvidia cards if these performance claims are accurate. Plus I'd guess Nvidia will launch a 3080Ti at ~$1000 sometime around the release of Big Navi.
I've heard they don't have enough production reserved at TSMC to make consumer GPUs on TSMC 7nm and they are only going to use TSMC for Quadro. Plus Samsung is cheaper than TSMC.
It’s interesting that the 3090 is a replacement for the Titan. The $1500 price tag is a bit higher than expected but considering the Titan cost $2500 it doesn’t seem too unreasonable.
tbh he pricing is much better than I expected. I'll get one for my new build, but probably a 3rd party as I'm not convinced by the cooling on the FE as of yet.
Interesting that this card has 8K capable HDMI 2.1 but not DisplayPort 2.0. Wonder when we start seeing DP 2.0 support in products, VESA said late 2020 in the press release.
Apologies if this question is super naive, but is there a good reason for ongoing development of both HDMI and DP? At least for my use cases (home entertainment, and work computers) both seem roughly equivalent.
The devices I've bought recently have tended to support HDMI more than DP. So I got the impression that HDMI was "winning" and DP would fade away.
But now it seems like vendors are moving towards video-over-USB-C cables. And the "Alternate Mode protocol support matrix for USB-C cables and adapters" table in this article [0] seems to indicate that USB-C cables have broader support for DP than HDMI. Which makes me wonder if vendors will converge on DP-protocol-over-USB-C-cable?
This makes me nostalgic for the relative simplicity of DVI.
HDMI is all about DRM. USB-C is just one of many reasons DP hangs on. The ability to chain monitors is huge for digital signage and other display uses too. Let's hope both continue to be developed since for computing display port is far more useful and free of at least some of the DRM hell of HDMI.
However, Wikipedia says it’s capable of 8k with a proprietary lossy video codec called Display Stream Compression:
DSC is a "visually lossless" encoding technique with up to a 3:1 compression ratio. Using DSC with HBR3 transmission rates, DisplayPort 1.4 can support 8K UHD (7680 × 4320) at 60 Hz
As much as I like to hate Nvidia for all of the right reasons, this is pretty big and might make me compromise my morals until AMD comes out with something that can compete.
I find the naming conventions for computer parts utterly confusing. I’m looking to step away from Apple and do my first (AMD) PC build. Need to find a good overview to read through.
Why do I need Wayland? Xorg works just fine. Actually better than fine, since you can actually run a wide variety of window managers without having to resort to xwayland
xorg is a security nightmare. Also, wayland itself works as well (even better performance wise), it just takes time for all the applications to catch up and enable it as a backend.
All these crazy graphics cards, they still couldn't figure out high density VR displays. I want amazing VR with super clarity. Then I'd invest whatever money they want for a graphics card. LCD gaming just doesn't cut it.
It's not that they can't figure out high density VR displays, it's that they're prohibitively expensive to produce. Display miniturisation is not a problem domain that a lot of tech is focused on, so progress is necessarily slower than the more profitable areas.
The pricing seems very good. Our company write a lot of CUDA code, mostly for real-time audio processing. It's amazing how much performance you can get with a desktop PC these days. These really are Supercomputers on a card.
These look great! It’s amazing how much better hardware gets annually. The only thing I was hoping for that wasn’t mentioned was hardware-accelerated VP9 encoding, but we can’t get everything we want in life.
If we were to imagine that VP9 hardware encoding by Nvidia would hold the same standard as their H.265 hardware encoding then we can stop holding our breaths, as we have not missed out on anything of any value what so ever.
For H.265 their encoder is fast, yes, but the quality per bitrate is complete rubbish, requiring higher bitrate than a good H.264 encode yet still contrives the gruesome trick of looking far worse, which entirely offsets all and any point with H.265.
No. Turing's NVEnc for H.265 is great for speed (as usual) and good for quality with Bframes support when compared to other HW encoders. (Ice Lake's encoder is also good but it's released later). It shouldn't be complete rubbish. IMO it's usable not only for streaming but also for archiving. Don't compare quality with x265, that should be good quality and slow.
Oh interesting, I didn't know the Nvidia encoder was regarded as trash. What tools can one I use to evaluate visual quality of a video file? My proxy for quality has always been bitrate, but I know that bitrate is just chasing visuak quality anyways...
It's worrying to see GPU cooling system dumping the heat directly onto the CPU, RAM and motherboard components. That's the only thing that left me skeptical after watching the presentation.
A few weeks/months after release, 3rd parties are going to come out with a bazillion different thermal solutions, including self contained liquid cooling, like they always do. So that's a minor issue. Just don't buy the reference model.
What do you mean? The fan design is pretty standard. Either way it won’t matter, as hot air should be extracted from the case, it’s not gonna meaningfully change the temp on anything it hits.
The back fan flows the air through a hole in the GPU, past the thermal tubes and carries the hot air above the GPU to the upper part of the motherboard. Which is not in any way a standard design with the hot air going directly outside the PC case.
The non standard part is pushing hot air out of the back of the gpu itself - “dumping the heat directly onto the CPU, RAM and motherboard components” to be extracted by the system fan is what every existing design does...
I'm glad the 3080 is launching first so I can try getting one of those and if they are sold out I have time to think about whether I want to try dropping more than twice as much on a 3090 instead.
Question: what is the sentiment around how this may affect pricing of the non-consumer cards? There was already a disparity, but the gap is now indecent - something like a T4 looks to be worse than a 3090 and more than double the price on basic specs (understanding there are some elements designed specifically for server environments). Is there any potential nVidia will have to shift on these prices or do we have to wait for a next generation of those before we can expect more affordable data center GPUs?
I got a 2080-Ti last year and I have yet to really run into a situation where I couldn't play a modern game, at 4K, with almost everything set to max (but I also don't game over 60fps).
I'm sure this will really push those 4k/120Hz displays, but I doubt the average/causal gamer will really care about this series for a few years.
From a fellow 2080 Ti owner this kind of statement that you can max things out without issue is usually true with the latest flagship but falls apart once the next generation comes out and provides a higher max performance envelope for settings to target.
Expect the games demoed in the event due to come out soon (cyperpunk, cod, etc) to not play as well when maxed out as those that were released while the 2080 Ti was the flagship.
I am curious how quickly Nvidia will add these to their GeForce Now streaming servers. As of right now, they only stream in 1080p and it seems this could allow them to stream 4k for about the same hardware cost. I'm personally not in the market for a gaming desktop, but happily subscribe to GPU as a service.
GFN has been a bear, very hit or miss. Current generation of hardware is either 1080 or 2060. If they start adding 30xx (which they have hinted at in this past) that would be great.
The FX series was pretty bad and around the same time AMD's X cards (eg X700) would burn out within a year if you gamed on them, but not fail. The FPS would slowly lower over time. This was due to a lack of a fan inside of the case.
Ah. It seems unreasonable to me to consider the 2080 and 2080Ti as different generations when they were released at the same time. Also I think we should factor that the 3080 at $700 costs _half_ what I paid for a 2080Ti in December (The ASUS RoG Strix, a premium 3-fan model, was going for like $1400 before it was discontinued). By NVidia prices, the 3090 is the true price successor to the 2080Ti.
You're right, I shouldn't likely be distinguishing the Ti series (although it is usually a 'toc' generation, like the just announced gen, as Ti variants were not announced).
But that does leave the 16xx generation which was released wholly on its own, in its own year.
Comparing 3080/3070, 3080 is 146% greater TFLOPS than 3070 and VRAM capacity and speed is different. Meanwhile 2080/2070 was 135% greater TFLOPS in same MSRP with 3080/3070. 3080 looks very competitive.
Sure, if the bold 2x performance increase claim is to be believed. I'd wait for benchmarks to validate that.
Plus for me, Nvidia is simply DOA on Linux, due them refusing to upstream their driver and hindering Nouveau to reclock properly. So even if AMD won't outdo them, I still won't touch Nvidia.
Depends, it works for cases Nvidia care about. But what you call integration means all other cases :) And there it simply falls apart or takes decades to be fixed.
Ditto. But it’s pretty much certain that even if the new AMD cards don’t match the top Nvidia performance, they’re still going to be competitive or better on cost/performance. So you don’t have to worry about missing anything if you go AMD for Linux.
I’m just glad that we’ve finally gotten past that ceiling at ~13 TFLOPs. Nvidia has been hobbling along for a few years, so a breakthrough is nice.
This looks reassuring. After the first couple rumors/teasers, especially regarding power consumption, I feared that NVIDIA mostly just sat on their hands and would just release something that's mostly a bigger version (more execution units, ram) of the current Gen. I think they did that once some years ago. Seems they actually did improve on the technical level too for 30xx. :-)
These numbers never meant anything. I run a 1050Ti on a 200W PSU - nvidia recommends 450W minimum. Add the TDP of your GPU, CPU and add about 100W for accessories = what you actually need. Nvidia recommends a much more powerful PSU than needed just in case.
> Nvidia recommends a much more powerful PSU than needed just in case.
The two things to watch out for here are
1) Cheaper PSUs can't always actually hit their claimed wattage, particularly not in real-world heat scenarios
2) CPU & GPU both use the 12V rail for their power, and not all PSUs can deliver all the rated wattage on the 12V rail.
Any decent to good PSU won't have either of those issues, most list their rated wattage entirely on the 12V rails these days.
So for example let's assume 250w for the GPU average and 120w for the CPU average (turbo & boost & all that). A 400w PSU could technically do that, particularly since if your only drive is an SSD your "accessories" are basically a rounding error. But if we take this 400W PSU for example: https://www.newegg.com/coolmax-i-400-400w/p/N82E16817159140 it can only deliver 300W on the 12V rail. Not enough. By comparison this EVGA 450W PSU can do a full 450W on the 12V rail alone: https://www.evga.com/products/product.aspx?pn=100-BR-0450-K1
That's a 150W useful difference in this scenario even though the "rated" power only differs by 50W.
I feel like this used to be more true in the past. These days CPUs can exceed their official TDP by quite a large margin, and while in theory they should only do so temporarily, many motherboards default to unlimited boost clocks. (Then again, perhaps you're never going to fully utilize both CPU and GPU at the same time...)
That is correct. Nowadays a 75W TDP Intel CPU can use as much as 200W for short bursts. That wasn't the case in the past. However, it should still be possible to find out that maximum draw value for many motherboards and pick a PSU accordingly.
Yea I think if you run the calculations and you have room you should be good if your PSU is just slightly above the power requirements. If you run an unusual amount of memory or HDDs/etc, you might want to calculate everything manually rather than just assume its 100 watts though.
I have never had a PSU fail, but supposedly unlike pretty much every other component if it fails its possible it will destroy your GPU/CPU/MB so it makes sense to spend a little extra on a good PSU.
Probably my best component purchase ever was a 1050W Modular PSU in 2014. It was an old model even then and apparently no one wanted 1000W+ power supplies back then because it was on clearance. It should still be good for a 3090 and probably even a 5090 when I upgrade again in the future.
This is likely a cautious recommendation on nVidia's part.
A 2080 Ti + 8700k system used 450W (nvidia recommended a 650w psu). While high end CPUs have gotten a bit more power hungry with higher core counts on the 10900k/10700k/3900x/3950x, I'd be shocked if a 650W PSU couldn't handle a mainstream CPU + 3080.
nVidia's recommendation is based on "we don't want people pissed off because they put it in a system with a 3990wx at the recommended PSU capacity and it didn't work"
You can see on Nvidias product page they are calculating based on a 125W processor. My 3700X is 65W so I'll have enough room for a 3070 easily on a 600W PSU.
Just to clarify: Of course a 65W CPU will still consume more power due to bursts and boosts but as far as I know AMD has a hard limit of about 90W for their 65W-rated CPUs, which is reassuring. I‘m glad I opted for the 3700X during my last upgrade - very efficient CPU indeed.
As someone who runs data models at home in addition to 3D rendering, 3090 is a must buy for me. I imagine it will be sold out within minutes and supply will be an issue for months.
Replying to myself. Still not sure about the form factor, but, unfortunately, it looks like the RTX 30 Series cards - well, at least, RTX 3090 - have a different socket interface when compared to 2080 Ti (12 pins versus 8+6 pins).
Higher resolutions IMO are not about the models themselves, but the edges of them.
A character model can look decent in 1080p, but the edges of the model in front of the background will be jaggy. Various anti-aliasing techniques can only do so much.
Besides, I know I'm more interested in higher frame rates. I'd rather do 1440p @ 144 hz than 4K @ 60 hz.
And I don't think there are even any 8K monitors out yet.
So it seems that the 3090 will be priced at 1499$. This is kind of insane.
EDIT: For people comparing this to the Titan RTX, no. This GA102, not GA100. It's the cut-down version of Ampere. GA100 will come out, and it will be even more expensive.
Virtually nobody needs a 3090, much less a gamer (let's be honest, though, many will buy one regardless). For the people that actually do need that horsepower, it's unbelievably cheap for what you are getting. You could have easily paid twice that a year ago for less.
> Virtually nobody needs a 3090, much less a gamer
I'm tempted to get one just to avoid having to think about upgrading a graphics card for another 10 years. Plus I can do some ML messing about as well for resume-driven development.
Best I can get for $500 now (~£377) is an 8GB GTX2060 - 5x fewer cores, 1/3rd the RAM, 2/3rd the memory bandwidth of the GTX3090 which is £1399. Plus I really don't want to upgrade my PC again for at least 5 years - just done that and been reminded of why I hate it.
> Virtually nobody needs a 3090, much less a gamer (let's be honest, though, many will buy one regardless).
That statement is absolutely true regarding me. Honestly, I don't need a gaming system at all, let alone one with such a powerful GPU. But even in terms of my leisure-time gaming, I could never justify the price difference over a much cheaper card.
Still, I could imagine it making a lot more sense for other people. E.g., pro gamers, people with big entertainment budgets, or people using CUDA for number-crunching.
Well yeah, that's just how it is because of progress. It's still more expensive than last year's XX102 chip. That said, it's only about 50% faster than a 2080Ti at everything except ray-tracing. 50% faster, 40% more expensive.
Back in Kepler, a 780Ti was 800$, and it had the GK110 chip, which was the full-fat chip. Now, the cut-down chip costs twice as much.
Yeah. Even if it is not a Titan, it is at a price that I am glad to pay for the performance (if the 1.5x faster than Titan claim holds true). Bert / Transformer models are incredibly memory hungry, and a sub-2k graphics card with 24GB memory is great to have. Also, its number of CUDA cores seems to be slightly more than A100, would be interesting to see benchmarks once it comes out.
Not to mention that it still has NVLink support! The 3-slot design is a bit challenging and for 4-card workstation, I need to rework my radiator mount to make space for another PSU.
Well, it's not just 1.5x faster than the Titan, it really is 1.5x faster than the aftermarket 2080Tis too since they basically are the same chip anyways.
If you're happy paying 40% more for a 50% faster card, that's okay. I just don't think it's very good for the industry.
> EDIT: For people comparing this to the Titan RTX, no. This GA102, not GA100. It's the cut-down version of Ampere. GA100 will come out, and it will be even more expensive.
That doesn't mean it's not the Titan equivalent for this generation. Titan X(pascal) and Titan XP were both GP102, and Titan RTX was TU102. AFAIK, only Titan V used the "100" chip, and that was sorta a fluke because there was no smaller volta chip. (and 3090 was explicitly introduced as the Titan RTX replacement)
Actually, the Titan, Titan Z and Titan Black as well as the Titan V and Titan X. Titans XP, Xp and RTX were basically just overclocked 80Ti chips, sometimes slightly unlocked. They are the outliers, and widely regarded as a scam, locking away 60$ of memory behind 1500$, with a bit more CUs yet about as fast or slower in aggregate than aftermarket 80Ti models.
It's not the Titan, because it's not the biggest chip, and also, it's not called "Titan". It fits the motif of the 2080Ti almost to a T.
If you want 144hz, even a RTX 2080 Ti can't hit it on 1440p on max with most games. I have a 2080S and I'm pretty disappointed with the performance, especially with Ray tracing on.
For example, RDR2 runs at around 50-60fps on ultra on my rig. Very disappointing.
Unfortunately not since Nvidia has had such a monopoly on the high-end GPU market.
Interestingly enough, it looks like the $379 in 2015 dollars worth $415 in today's dollars which makes the $500 for the 3070 seem slightly less shitty. I didn't expect the cumulative inflation to be 9% since then..
It's long overdue to have widely available metrics for latency in consumer tech. Many hardware/software setups have absurdly long latency for no good reason other than that it's difficult to measure. People underestimate the subconscious effects it has on your usage of technology.
I couldn't be happier that phone and monitor manufacturers are finally starting to compete on refresh rates >60 Hz. It's far more important than wide gamut or 8K.