This is inference only. AMD should invest into the full AI stack starting from training. For this they need a product comparable to NVIDIA 4090, so that entry level researchers could use their hardware. Honestly, I don't know why AMD aren't doing that already, they are best positioned to do that in the industry landscape.
The hardware is not their major problem. They have been failing super hard at the software side of machine learning for a solid decade now.
It seems like pure management incompetence to me. They need to invest a whole lot more in software, integrating their stuff directly into pytorch/TF/XLA/etc and making sure it works on consumer cards too. The investment would be paid back tenfold. The market is crying out for more competition for Nvidia and there's huge money to be made on the datacenter side but it all needs to work on the consumer side too.
AMD has finite resources, like any company, and they’ve been focusing on CPU/datacenter dominance, which to me is both the safer bet and the more-lucrative bet. It wasn’t that long ago that AMD was on the brink of bankruptcy (~2016), so I appreciate that they’re not trying to divide their attention.
Their attempts at entering the ML space so far have been failures, and they are wise to hold off on really competing with Nvidia until they have the bandwidth to go “all in”. Consciously NOT trying to compete with Nvidia is the reason they didn’t go bankrupt. Their Radeon division minted from 2016-2020 because they focused on a niche Nvidia was neglecting- low-end/eSports (also leveraging their APU expertise to win PS4/Xbox contracts).
I think Nvidia will eventually lose its monopoly on ML/AI stuff as AMD, Apple, Qualcomm, Amazon and Google chip away at their “moat” with their own accelerators/NPUs. As mentioned though, the Nvidia Edge really comes from CUDA and other software, not the hardware. I doubt that Apple, Qualcomm, Amazon or Google will be interested in selling hardware direct to consumers. They want that sweet, sweet cloud money and/or competitive advantages in their phones (like photo processing). I don’t want to be paying AWS $100/mo for a GPU I could pay $600 once for. I do think AMD/RTG will go hard on Nvidia eventually, and it will not matter whether you have an AMD or Nvidia GPU for Tensorflow or spaCy or whatever else.
> AMD has finite resources, like any company, and they’ve been focusing on CPU/datacenter dominance, which to me is both the safer bet and the more-lucrative bet.
Cloud/datacenter based ML is a huge growth market. Having the same software work on a consumer GPUs and enterprise ML cards is one of Nvidia's competitive advantages.
Well they definitely can, and do. They have made the decision that the reward isn't worth the risk.
The idea (not aiming at you here, you didn't say this) that senior leadership at AMD is unaware of NVIDIA's lead in this space, and haven't repeatedly considered whether to invest in competing, is absurd. Likewise the idea that anyone outside of AMD understands better than AMD does what it would take in terms of investment _and opportunity cost_, is also absurd.
Senior leadership at AMD isn't dumb. The fact that they're not doing something we want doesn't make them dumb, either. Again, not aiming at you with this little rant :)
They may not be dumb, but they may have narrow vision (which I suppose is still dumb). Bad decisions in top tiers of the largest world companies are not unheard of.
> The idea that senior leadership at AMD is unaware.. Senior leadership at AMD isn't dumb.
Lets try this with another company:
The idea that leadership at Lehmon Brothers is unaware of the fact that they are trading subprime loans is absurd! The leadership isnt dumb
The Idea that leadership at Being is unaware of safety issues with 737 Max is absurd! How could you suggest that anyone outside boesing understands better than they do the risks involved?
The fact that a couple of other companies have had dumb leadership in no way proves that AMD's is dumb. You're essentially claiming that because Lehman and Boeing had dumb leadership, all senior leadership at all companies is dumb (because there's nothing linking AMD to these other companies except that they're all companies). And that is an absurd claim.
The idea that AMD management can't possibly have made any bad decisions is the absurd thing here. It's entirely possible that AMD carefully considered Nvidia's position, carefully considered their strategy, and confidently made the wrong decision. It happens all the time in all sorts of companies.
I think it's very clear with the benefit of hindsight that not investing enough into the software side of deep learning early on was a bad decision. But it was obvious to me even at the time and I said as much to anyone who would listen (e.g. seven years ago https://news.ycombinator.com/item?id=12258027)
Last I checked they see deep learning training as a niche market, their strategy is to try to win big contracts (HPC etc) and then supply software specifically for that. Then "the community" will supply software. Having spent a bunch of time beating my head on this and related walls it's not clear to me that they're entirely wrong from an economic standpoint. Remember that 2/3 public cloud providers have their own chips as well as NVIDIA's so it would be tough to negotiate a good deal. As a user it's super irritating to be stuck on NVIDIA especially when Jensen gets up on stage to say "haha, Moore's law is over, stop expecting our products to get cheaper."
I hope they change their minds. At least now that generative models are becoming somewhat popular. I'd love to be able to get an AMD card to run generative models, but to the best of my knowledge, they only run on Nvidia hardware
I wouldn't hold my breath, and anyway at this point NVIDIA has faster chips and more supported software all the way down the stack. My previous startup tried to solve some of these problems and we built what is as far as I know still the only reasonably complete device-portable deep learning framework. Today something like an RTX 3070 is a good budget option for small experiments and you can always lean on a cloud provider if you need more compute temporarily. Hard to beat a TPU pod when you're in a hurry.
no, they need a product good at training and gpu compute at a
reasonable price
that product doesn't need to be good at rendering, ray tracing
and similar
sure students and some independent contractors probably love getting both a good graphic card and a CUDA card in one and it makes it easier for people to experiment with it but company PCs normally ban playing games on company PCs and the overlap of "needing max GPU compute" and "needing complicated 3D rendering tasks" is limited.
through having 1 product instead of two does make supply chain and pricing easier
but then 4090 is by now in a price range where students are unlikely to afford it and people will think twice about buying it just to play around with GPU compute.
So e.g. the 7900XTX having somewhat comparable GPU compute usability then a 4080 would have been good enough for the non company use case, where a dedicated compute-per-money cheaper GPU compute only card would be preferable for the company use case I think.
Long time ML worker here. People work in one of 3 ways:
1) Consumer Nvidia GPU cards on custom PCs
2) Self hosted shared server
3) Cloud infrastructure.
There is no "GPU compute only card" that is widely used outside servers.
> company PCs normally ban playing games on company PCs and the overlap of "needing max GPU compute" and "needing complicated 3D rendering tasks" is limited.
The "don't play games thing" isn't a factor. Most companies just buy a 4090 or whatever, and if they have to tell staff not to play games, they say "don't play games". Fortnight runs just fine on pretty much anything anyway.
my point is a card bought for doing GPU computer not being able to work as a normal graphic card is not a problem
I'm aware that currently GPU compute only cards are not widly used outside of the server space.
But that's not because people need the consumer GPU features (bedsides video decode) but because the economics of availability and cost lead to consumer GPUs being the best option (and this economic effects don't just apply to customers but also Nvidia itself).
Because entry level researchers shape the industry in the long term. I'm in academia, I worked at two universities and I have not seen a lab that uses non-NVIDIA hardware for research. Majority of graduates go to work in the industry.
I used to believe this idea that the tools in academia would carry over to industry. Now I think it's only weakly true, that is there's not really a big barrier to switching to other options like TPU or maybe Trainium (haven't tried it). Supporting independent researchers gives you the Heroku problem, they may like your product but as they get more sophisticated and go to production they'll accept significant pain to save money, scale better, etc. You basically have to re-win that business from scratch and the technical tradeoffs are very different at that point.
MI200, MI250, and MI300 should work for training. They don't have an exact equivalent to the 4090 but that may be setting the bar too high. Nobody can deliver everything that Nvidia has but better.
It will be harder for a small academic lab to afford MI300, whereas everyone can purchase a few cards costing $1500. And even if I had money to buy MI300, I wouldn't - it is a too risky investment, because we have no idea how suitable they are for common AI research workflows. They need to lower the entry bar, so that people can try out their hardware. Even 80% of the performance of 4090 would be enough at an appropriate price point.
4090 (or 3090, 1080Ti and so on) is a high-end consumer GPU, but at the same time it is an entry level GPU for AI researchers. Don't forget that workstation cards (RTX 8000) let alone server-grade GPUs such as A100 are an order of magnitude more expensive.
Chances of you publishing something in ML improve proportionally to the amount of hardware you have access to. Or to put it another way, the less hardware you have, the smarter you have to be to publish something in ML.
IDK, I want to train up some AI and have the choice of using google colab or buying a GPU that can do it within a reasonable timeframe.
Not going to be spending $10k like the tortoise-tts guy (was looking into that project last night) but $2k might be doable for a hobby project. Plus I’d have a computer at the end.
Are these programmable by the end-user? The "software programmability" section describes "Vitis AI" frameworks supported. But can we write our own software on these?
Is this card FPGA-based?
EDIT: [1] more info on the AI-engine tiles: scalar cores + "adaptable hardware (FPGA?)" + {AI+DSP}.
It's very likely FPGA-based; Xilinx is an FPGA company. This is being pitched as an "AI accelerator", but "Alveo" as a product line existed before AMD's acquisition of Xilinx, and other "Alveo" products exist (https://www.xilinx.com/products/boards-and-kits/alveo.html) that are marketed for other purposes, while really just being Xilinx FPGAs pre-programmed to perform specific other tasks, with some domain-specific DSPs + interconnects around the edges.
It's possible that AMD could have reworked an existing Xilinx design to incorporate RDNA chiplets in place of some of the FPGA-gate-grid chiplets, creating a heterogeneous mesh; but I find it just as likely that AMD just took their VLSI for an RDNA core and loaded it onto the existing FPGA.
It's not a traditional FPGA chip (lots of luts and flip flops). The "AI Engine" is basically hardened chiplets that are working alongside soft logic chiplets and I/O. This is how they're able to get their performance/power numbers
I suspect that it still has some fpga fabric attacched to the ai engines. The two parts are separate, but according to Xilinx docs (talking about Versal Soc), they are supposed to work togheter
Interesting. It seems then that the xdna architecture in the Ryzen 4070 is nothing more than a port of the existing Xilinx Versal cores (fpga+ai engine)
If this is based on fpga tech (xilinx) I don't think it will have a cost/benefit edge over asics. Why not do their own TPU like Google did? Nowadays even embedded MCUs come with AI accelerators (last I heard was 5TOPS in a banana pi-cm4 board - that is sufficient for object detection stuff and perhaps even more).
Probably because it's a product aimed at datacenters and cloud providers who work directly with Xilinx/AMD to develop it, so they already know the price.
Maybe. Running inference at 10 fps is probably plenty. But that doesn't mean you only have to do 10 fps of H.264/H.265 decoding. I think the most common scenario is for the input video to be e.g. 30 fps with mostly P frames that each depend on the prior frame in a chain. In that case, you need to decode almost [1] 30 fps to get 10 fps of evenly spaced frames to process.
[1] You could skip the last P frame before an IDR frame, but that doesn't buy you much.
Still depends. As it happens, I'm developing my own open source NVR software, [1] so I know a bit about this. Some cameras are fairly good about this, supporting the following features:
* "Temporal SVC", in which the frame dependencies are structured so you can discard down to 1/2 or 1/4th of the nominal frame rate and still decode the remainder.
* Three output streams, which you could configure for say forensics (high-bandwidth/high-resolution/high-fps), inference (mid-bandwidth/mid-resolution/low-fps), and viewing multiple streams / over mobile networks (low-bandwidth/low-resolution/mid-fps).
* On-camera ML tasks too. (Although I haven't seen one that lets you upload your own model.)
But other cameras are less good. E.g. some Reolinks [2] only support two streams, and the "sub" stream is fixed at 640x352, which is uncomfortably low. Your inference network may not take more resolution than that, but even if not, you might want to crop down to the area of interest (where there's motion and/or where the user has configured an alert) to improve quality. (You probably wouldn't pair that cheap Reolink camera with this expensive inference card, but the point stands in general.)
Even the "better" cameras' timestamp handling is awful, so it's hard to reliably match up the main stream, sub stream, analytics output, and wall clock time. Given that limitation it'd be desirable to just use the main stream for everything but the on-NVR transcoding's likely unaffordable.
Price wise if this card is $5,000 that's $52 per where you don't need any onboard smarts handled by the camera in a space where commercial cameras are hundreds of dollars to buy or replace to have the particular smarts you're looking for that day. I've done a few PoCs in the smart city/smart retail space they are advertising here and they pretty much end up falling into the "everything must be pre-processed as much as possible and sent to the cloud" or "everything must be dumb and sent to the central recorder" buckets as anything in the middle creates a bad cost balance where you're neither optimising hardware+simplicity costs or data+cloud costs. I'll admit though I don't normally go out to sell cameras all day it's just something we've added as clients in part of a larger connectivity rework (CBRS/LTE/Wi-Fi/GPON/traditional wired) and we typically partner up with some specialized company on the video processing use case. The onboard camera processing is usually about justifying a cloud pitch ("we use data to send video when something interesting happens" or "we send only the best picture of the face in HD to save bandwidth but still be able to ID them later") not so much letting you go in and solve your own problem. One exception I ran into was license plates at a car wash outfit where they were able to send the plate numbers back to their main app but that probably came from being a pre-baked solution for road tolls.
I also have a sneaking suspicion using lower channel counts let you raise the FPS but the max of 96 channels is the hard limit, tuned to allow up to use cases like recognition from unprocessed feeds but the documentation access seems to be a manual approval process so I can't verify for sure.
> Price wise if this card is $5,000 that's $52 per where you don't need any onboard smarts handled by the camera in a space where commercial cameras are hundreds of dollars to buy or replace to have the particular smarts you're looking for that day.
Good point. At that scale, the price might make sense. (I'd still hesitate to buy this card, though. Based on experience with Amazon VT1 instances, I don't have any faith in Xilinx's software quality.)
There are much lower-cost solutions if you don't need that many cameras, e.g.:
* The Coral TPU is nice and cheap. I keep hoping to see a new version and/or someone making M.2/PCIe cards with several of these chips on it. It doesn't do the video decoding, though, so you need other hardware for that.
* There was an Axelera card just announced. [1] I'm curious to read the reviews when it actually ships to folks.
* The newer Rockchip SoCs advertise decent video decoding and some ML acceleration. I have one and will be trying it out sooner or later.
> The onboard camera processing is usually about justifying a cloud pitch ("we use data to send video when something interesting happens" or "we send only the best picture of the face in HD to save bandwidth but still be able to ID them later") not so much letting you go in and solve your own problem.
My software's more aimed at the home/hobbyist side of things. There some folks go with the canned/cloud stuff (Ring/Nest/whatever) similar to what you're saying. Some do everything at home with e.g. BlueIris and use the on-camera ML stuff as it is. The lack of flexibility (mostly due to closed-source, low-quality software IMHO) is a real problem though. Some folks use something like Frigate that does on-NVR analytics, and I'll eventually add that feature to my own software.
> I also have a sneaking suspicion using lower channel counts let you raise the FPS but the max of 96 channels is the hard limit, tuned to allow up to use cases like recognition from unprocessed feeds but the documentation access seems to be a manual approval process so I can't verify for sure.
I assume it's int8 operations (FMA would count as 2). At least that's the case with FLOPs, TOPs for AI accelerators is basically the same measure, with the changed acronym reflecting the non float data format.
>> I have no problem imagining a security camera application needing to monitor quite a few video channels.
As a joke I sometimes tell people the automatic flushing toilets in public bathrooms work by having a little camera monitored by someone in a 3rd world country who remotely flushes as needed, while monitoring a whole lot of video feeds. They usually don't buy it, but will often acknowledge that our world is uncomfortably close to having stuff like become reality.
On the inference accelerator? IIUC, the RAM is just to hold the model and whatever state it needs during a particular inference operation. I'm not an expert on ML but AFAIK 16 GiB is plenty. I suppose it'd also need to hold onto reference frames for the video decoding, but at 1080p with e.g. YUV420 (12 bits per pixel), you can hold a lot of those in 16 GiB. edit: e.g., 4 references for each of the 96 streams would take ~1 GiB.
Even on the host, 16 GiB is fine for say an NVR. They don't need to keep a lot of state in RAM (or for that matter to do a lot of on-CPU computation either). I can run an 8-stream NVR on a Raspberry Pi 2 without on-NVR analytics. That's about its limit because the network and disk are on the same USB2 bus, but there's spare CPU and RAM.
I have models in production that currently monitor ~400 cameras with an addition of 2-3 cameras/month. If it were cheap enough, it would be useful for our use case (Quality Control). We generally pull from cameras roughly 6400 pixels per region of interest, of which one instance may have 4-30 RoIs across N cameras.
Not sure about open information about the models, but from tooling/infra we are running k8s w/ in-house API for image acquisition. Features are defined as x,y coordinates denoting center of a feature, with a pixel count denoting size of rectangle in each direction from center.
I won't even take a look at the numbers unless they show a PyTorch model running on it, the problem is the big disconnect between HW and SW, realistically, have you ever seen any off-the shelf model running on something other than NVidia?
It's for inference only not training. In this use case, there is lots of device that's not Nvidia. For server you have Google tpu, for more close to public there is the Apple Neural Engine for example.
