At $60 - its able to run anything x86, plus hardware transcoding and a cpu about 3x more powerful than a RPI 5, not sure why anyone would even want a raspberry pi. Not to mention the built in 2.5G nic with PoE and m.2 slot. Adding a PoE and m.2 hats on a raspberry pi would add another $25-40 to the price. Hope they ramp up production on these Radxa boards.
Edit: another source with better image of the form factor. It might actually fit a raspberry case.
i've done the rant many times before [1] but frankly rpi has almost never made sense. there were always things like the ECS Liva/Liva X for $100-125 that were fully-featured, for roughly the same price as a rpi once you consider all the shit you'll need to get that $45 computer booted up. The Liva (and similar booksizes/net-tops, and similar things like AM1 platform from AMD) benefited hugely from using standard drivers and standard BIOS versus a massive amount of early ARM jank (the early RPi days a lot of things were actually soft-float, let alone niceties like UEFI!).
the best argument for rpi has always been the combination of GPIO and linux. But that's a double-edged sword too, because linux isn't really very good at hard-realtime. And you can always just get a Bus Pirate or similar tool for doing just plain GPIO. And nowadays the ESP32 has basically completely displaced it for a lot of those sorts of "GPIO glue" roles.
If you didn't need that, like if you just wanted a low-power fileserver or HTPC... the x86 stuff was better, because it actually worked reliably, at roughly the same price. Like literally just go buy an off-lease mini-tower or USFF mini-pc for $10 from a surplus store, even.
This goes doubly when you consider all the problems with the early rpi. Like basically it uses USB 2.0 (unidirectional 500mbps) as a system bus... but it also dropped USB packets under load due to bugs in the firmware[0], that nobody outside the Pi foundation could address because Broadcom didn't release the documentation for like 5 years.
What I've ways felt was a major advantage for Raspberry Pi compared to all of the other SBCs is that the Pi actually got regular updates. Maybe other SBC manufacturers started doing that too since I first got my hands on a Pi, but it definitely didn't use to be the case.
The ECS Liva would cost me €200 (€278 when avoiding webshops that look like scams) without memory. An RPi5B is €88. I don't think I'll get more than twice the use out of an ECS Liva if all I want to run is Pihole.
And remember, the point of the Raspberry Pi was not to bring a cheap, high performance computer to prototypers. It was created to aid schools in teaching kids about computers. They've changed their tune ever since just about everyone with a computer project started buying out their reserves, but making things easy for professionals was not part of the plan from the start.
If you have cheap access to better computers, make good use of that! I still find the RPi to be competitive for cheap, low-power compute, though.
> And remember, the point of the Raspberry Pi was not to bring a cheap, high performance computer to prototypers. It was created to aid schools in teaching kids about computers. They've changed their tune ever since just about everyone with a computer project started buying out their reserves, but making things easy for professionals was not part of the plan from the start.
I go into this in my previous comment that I linked above, but I think it was a shitty product for that niche too. The original non-microSD version had massive problems with the SD cards slowly softening and deforming enough to cause unreliable contact with the socket, which was one of the failure modes that caused corruption.
Another were the power brownouts, etc. People forget because usb chargers have gotten so much better over the years but 10 years ago a regular usb charger could generally not run a raspberry pi stable under load. Worse, it would appear to be stable and just silently corrupt the SD card due to voltage brownout (writes don't "stick"). Today we are seeing the same problem with the RPi 5s, where they have simply exceeded the capability of what can be delivered at 5V and aren't reliable.
You could not even PXE boot them without having a SD card installed to provide the blobs, so there was no getting around this at the time.
Pointing to the "you can pull the SD card and reflash it!" as being some kind of massive advantage is ignoring the massive downsides the SD card brought with it, again especially until they finally switched to microSD so the cards didn't melt and warp. I cannot fathom someone trying to run a computer lab on raspberry pis, that would have fucking sucked.
I partly agree, the RPi Zero W is my ideal small physical computer with a full OS. Running it only from the overlay filesystem (to keep those terrible SD cards from becoming corrupted). Pi-Hole I run from a docker container on a fanless NUC-a-like. But for anything that needs GPIO, the Zero W is great (nice that this Radxz has a RP2040 built in to cover that).
The ECS Liva X is 10+ years old, it’s a product I bought in 2014. You’re seeing elevated prices from scalpers who are selling stock of a product that was discontinued 8 years ago, buying a new-in-box GTX 760 is going to run you a premium as well.
At the time, I bought one Liva for $100 and then two Liva Xs for $125 each a year or so later. All of them were 2gb memory/32gb emmc models.
I like to use a single remote for both my TV and Kodi, and the raspberry pi has the CEC HDMI thing which makes it possibke. It's not that common in SBCs AFAIK.
Unfortunately it's not cheap, although the same functionality could be offered by a USB equipped small microcontroller "sniffing" 2 wires from a pass-through HDMI port.
RPi is the only ARM SBC that I've seen good broad support for. You can run their Linux distro as well as mainline distros. Every other SBC I've tried has required some boutique distro with custom drivers that weren't generally available or worse still, just didn't work.
I will say, when the scalping of RPi happened over the pandemic, a lot of the x86 mini-pcs became a much better option in terms of TCO. I invested a bit into DeskPi cases for 8gb RPi 4b models. In the end, a $150 or so N100 box is a much better deal, with the case, psu, memory and storage. But it, of course, depends on what you're doing or trying to accomplish.
I have a solution which was developed and tested by another person under rPi. I know it's working on exact model. All I need to do is to buy exactly the same model of rPi, install the same OS, and deploy that solution there, as many as I need.
You need to be very naive to think that replacing rPi with a completely different thing will not bring any problems. Hell, even changing the rPi model does!
Sure. And another thing that matters: the producer of a faulty USB cable that led us to waste several days of work. And another: shitty battery producer that led us waste several more days.
This "little computer land" is flaky and hard, and rPi matters because it gives at least some common ground. I wish it'd be more universal, but it's not.
It's not just them, the majority of SBC manufacturers are on these bizarre old Linux forks that have been patched all to hell and will never see mainline. Kind of same situation with most smartphones, too (as they are all the same SoCs).
There are some very noble projects (like PostmarketOS for phones, and Armbian for SBC) who try to be "mainline first" to keep these fascinating little devices going longer (and more reliably). They do an amazing job considering what they have to work with, but these are not big companies with a lot of resources behind them, so if you care about this sort of thing (not you specifically, that's for anyone reading) then consider perhaps lending a hand (or a few bucks)!
But x86 boards have vastly fewer kernel & driver support problems.
It's not even close, not even in the same order of magnitude.
The problem is not "sbc", the problem is "arm sbc"
Running on a random "unsupported" x86 board is no more of a problem than an unsupported laptop. There are sometimes no drivers for some peripheral but it's nothing like arm where you can't even boot without a dtd and most of the hardware has some freak custom interface instead of just pcie and usb and well known chips.
Every single arm board is like it's own whole different platform. A Bananna Pi Foo A and a Bannana Pi Foo A+ both with nominally the same cpu and specs might as well be a trs-80 and an appollo guidance computer when it comes to bringing them up.
IMO the logic is reversed. Wintel phenomenon created enormous incentives for everyone to follow IBM PC "spec". That was odd, and it's ARM SBC situation that is more normal.
Well, it's a very standardized platform that is readily available and guaranteed to be available for some time.
This is why some hardware projects reply on it and the intricacies of its GPIO timing.
I agree that if you want a small computer to run your home NAS on, then sure, there might be better alternatives.
But if you want to (say) plug one into your old Amiga (with a PiStorm) so that it emulates the CPU's bus timings and runs as an incredibly fast M68000 CPU, there's no way around the RPi (unless you want to do a lot of hardware development yourself... if that's even possible since on most of these Chinesium clones, the GPIO is hidden behind a ton of bridges, busses and/or MCUs which would kill the latency in those tightly constrained use-cases)
I assume nobody would bother porting the hardware-related code to the "Chinesium Clone of the Week" every few months
one big advantage pi had over minipc is a lot of the minipcs are in the 30 watt range while the early pis were single digit watts. that advantage has decayed over time though...
A good number of the mini PC's idle under 12W. My i5-6500t does. Yeah they ramp up under load to 30W, but a before rpi5 to start to ramp up like that would've been way more compute than an RPi could offer.
Also notably there are some low power options too. My first kube cluster was on these nice 2014 Acer Chromebox cxi's; they idled at 6w! Crazy slow 2957u Celeron processor, but I used it a node as my desktop for ~2 years (with 16GB ram). But it had gobs of USB bandwidth, mpcie wifi, sata, gigabit, and if i'd sprung for i3 models it probably would have been downright pleasant to use versus the 1.6GHz no turbo dual core I was getting.
Personally I've avoided the atom cores generally, but there's been a variety of decent low power offerings of big cores for a while. Heck, my Ultrabook tablet is ~6W when running, with screen on, with a i5-7200u. 15W TDP core with some additional TDP up available.
Point is, there's been a range of pretty low power x86 offerings. Just, not always super obvious.
the n100 mini pc i got here runs at around 3.5 watts while running homeassistant with a unifi network server while having around two dozen devices connected. under debian 12 minimal.
How are you measuring that? I've been scouring the web for low-idle-power builds and 3.5 watts is far below the best builds I've found (assuming that 3.5 watts is measured at-the-wall).
I've measured about 5w from the wall with Pentium J5005 based Dell thin client which has a tiny fan, one SSD, two RAM modules. No idea how N100 compares to J5005.
Just a cheap measurement adapter you plug between the wall and cable.
The nice thing about n100
Is its ability to disable the iGPU in the bios, saving you a little bit of power. But you have to reset the bios on the board to regain any output if you might need it later.
Unless I'm running on battery, that's a pretty abstract win to me in the US. My i5-6500t's 10W for all 8760 hours of the year is 87kWh; less would be nice but my power bill is already under $10/year at that rate.
It is great that something on par is so affordable & even lower power though! And what I bought was a value because it was used, where-as now you can buy new. It's felt like the atom lineup was kind of slow rolling things & there hasnt been much availability for cheap AMD systems. These low price low power e-cores systems have really changed the game, have expanded the value proposition a lot versus the used-x86 mini PC or io limited ARM market's long stalemate.
Sounds interesting, could you please share brand and model? Also, do you know how many SSDs/HDDs you can put inside? Could be a good replacement to have a low power NAS.
Interesting, but they can jump up to 50W or 100W depending on the model. Which such a big swing, you really need to consider your specific use case to see what will be the actual power consumption.
Apple's still the only company that matches other Arm boards/servers on efficiency (perf/W) + idle power draw. AMD has been doing a little better, Intel slightly more so.
I am keeping an eye out for a cheap M1 Mac Mini on my local Craigslist because once you yank it out of its huge metal case, which is like half PSU, the board itself is quite small. Apparently, the PSU only outputs 12W across the entire power connector, so easily powered by a barrel jack. I quite like the idea of playing around with macOS as a SBC OS, but even just designing a smaller case with sufficient cooling would be fun too.
no, the booksize pcs of the time were pretty comparable to what you see today in terms of power. rpi was a bit cheaper and lower-power... but also vastly slower. Like a factor of 4x if not more (10x?).
Remember: single-core, ARMv6Z 32b, 700 mhz in-order, using USB 2.0 as a system bus and booting over that same shared 500mbps-unidirectional system bus, with 256mb memory. Versus dual-core x86 64b 1580 MHz out-of-order with a proper SOC design, and also having dedicated USB 3.0 for peripheral connectivity, with 2GB memory.
For say $80-100 loaded on a rpi vs $100-125 on a Liva/Liva X it was a no-brainer even at the time, and the liva still sipped power (because it was using e-cores). The 1B+ model improved things a little (512mb) but still, not a good showing for a product that competed directly with these x86 nettop machines.
And in the grand scheme of things... 1w idle/5w load vs 3w idle/12W load is not anything you're really going to notice, especially given the massive increase in performance it afforded.
Rpi team had great marketing. But it wasn’t the only game in town even back then. Nettops existed for a long time. They just didn't have the full-court press in the glossies.
Yes, that's always been the marketing gimmick. $35, for something with no case, no storage, no power adapter (and off-the-shelf adapters were not suitable at the time), that had a critical flaw with its system bus/all USB devices for the first 2 years of its existence. That bug didn't get fixed until feb 2014, which is basically the same time the liva launched.
also, booksizes/nettops didn't materialize into existence in 2014 either. and you have to remember that the raspberry pi 2 didn't launch until 2015, so actually they were direct contemporaries for at least a year.
so if you wanted to run the early-adopter softfp distro on a $35[0] SBC in 2012 then sure, I agree, go for it. But there were turnkey solutions in the x86 space that weren't that much more expensive.
ECS Liva according to reviews from 2014 would cost me about $170~180. Would still require keyboard/mouse and something to display with.
Sure, it would have a case and WiFi in that price.
It would also miss GPIO header, which was major selling point of initial Raspberry Pi efforts (before it became so popular for other uses, when it was still targeting education).
A lot of then-common phone chargers worked. A random SD card wasn't that much of an issue. A kit with charger, NOOBS SDcard, and random assortment of extra bits to play with the electronics regularly could be acquired for ~100 USD if not less.
If I wanted to play with small-size BayTrail x86, Liva was fine choice. If I needed a small general purpose computer, too.
Not a very good choice if I wanted to graduate from Arduino towards somewhat more capable setup without requiring either expensive I/O boards or hacking an IO board out of Arduino or another MCU.
like idk what to say dude but MSRP reviews aren't always reflective of street prices back then, things used to go back on clearance sales before moore's law died.
yes, you absolutely could get x86 stuff that was price-competitive with the rpi once you loaded it up. the $35 price is, as I have stated, a red herring considering all the stuff that nobody had on-hand in 2012 like proper usb chargers that could run a rpi stable, plus the massive kernel problems involved for the first few years.
It's $100 loaded cost for a raspberry pi 1b or 1b+ that doesn't even work, because of kernel bugs I already linked, or just buy x86 for $75-180 and get something an order of magnitude faster that actually works and runs standard x86 distros and drivers. Stop comparing 1/3 of a pc to a out-of-the-box pc as if the price means anything at that point, this is getting to be bad-faith.
Especially when the product didn’t fucking work.
Like it’s the same damn story as ROCm, right? People want serious financial consideration as if they have delivered a working product, minus the delivering the working product. Come back in 5 years when your product works.
I will sell you a brick with a piece of paper that says “10,000 tflops” taped to it. The perf/$ will be out of this world, but there will be some “usability issues”. PM me for details.
yeah, even for GPIO interfacing, there are far cheaper options (ESP32) than the raspberry Pi out there. The Pi was an appealing piece of hardware about 10 years. But now, not sure it makes sense to use as a low cost server.
Never understood RPis, because as a Intel user, you can already build such a system yourself quite trivially by getting a NUC, making a case locally with a 3d printer, and then using ZFS or Btrfs for the computer's filesystem.
A also "Never understood why people go grocery shopping to make their own food for $25 when they can go to a restaurant for a $200 meal, or go through the gabage behind the restaurant to pick up some throwaway scraps for nearly free."
If you have the know-how and factory to do final assembly, it's at least one more step in a long supply chain you have mastered. Broadcom or no Broadcom.
"At least" is certainly right. You'd be adopting the lowest-margin, most competitive and lowest-hiring work from China. And you'd still be paying a price-premium for a dubious improvement in quality, if any improvement at all.
Eventually people have to realize that America cannot re-take the jobs we sent to Mexico and China. The modernization we enjoy today is entirely predicated on the abject abuse of people that are protected by inferior governments and weak economies. To import their jobs means to import their standard-of-living.
Same problem, just worse. If America can't source butt-cheap STM-32s then our efforts at assembling boards might as well go up in smoke. But that's easy enough... just bootstrap domestic refineries for every precursor, purchase a few billion dollars worth of ASML products, hire and train a few thousand workers while you build a dozen-odd factories, convince Texas Instruments to license ARM hardware officially at-cost and design an American-made core, validate and sample the new chip to your assemblers and pass the research/development costs onto the consumer with every device!
That's the problematic bit. Now you have an officially-licensed and tested STM-32 core that can serve as a drop-in replacement at 10,000% of the cost of an identical Chinese bootleg model. Part of the problem is that the US has to pay for ARM licenses that China refuses to acknowledge, but another part of the problem is that our supply chain's just not built to compete. We're a digital import-state.
This board needs some supporting hardware that RPi doesn’t - a fan and wifi antennae, if I remember. Radxa includes these with a case as well, so the added cost is trivial, but at that point the footprint’s more than doubled.
(That’s not to say that this isn’t preferable; just that it makes different trade offs that in some cases don’t work.)
I think the price points for Pi 5 are staying where they are ($60/80 was launch price for Pi 5 4/8GB, respectively).
The fact the boards have a 1 and 2 GB resistor option for memory means I'm guessing those were planned (or are planned still?) and would logically slot in at a $40 price point.
That would give back some of the price-advantage for many Pi use cases, since at least for me, about half the things I use Pis for, I could do within 1 (preferably 2) GB of RAM.
I think it's highly dependent on the workload under test. The N100 includes Quick Sync (and the Pi 5 does not) so any processing that includes video transcoding will be a lot faster on the N100.
True, but the fact a pi can even compete on paper in some areas with a modern intel chip is pretty impressive. 3x isn't the power difference for all workloads but probably is for common ones like transcode etc.
I think we often forget just how far the pi has come from the armhf days.
Cool to hear… especially in the power specs for that proc. If you don’t mind, What specs does the transcoding machine have and what’s an example of a workload it handles well?
It can transcode a 4K HDR10 movie encoded at 60Mbps down to 1080p SDR ~15Mbps at around 78fps. 4K HDR to 4K SDR at around 35fps I think. Those are typically the most intensive workloads I throw at it. Dolby Vision encoding is a bit harder on it, I think it uses OpenCL rather than VPP tonemapping for that, so those can drop as low as like 28fps when going 4k DV -> 4K SDR, but for my purposes that's the most intensive workload and those would all be 24fps movies, so it generally just works great.
I also have a fair bit of anime encoded in Hi10p h264 format, and it has no trouble decoding and re-encoding those (to h265) with burned-in subtitles. We're talking 90+fps. I note Hi10p mostly because hardware decoders for it don't exist.
It's at the encoding step that we usually see it getting close to its limits for transcode speed, hence why 4k HDR to 4k SDR is the most intensive load. It's probably also safe to assume that when transcoding down to 1080p Jellyfin puts the scaling down filter first in the ffmpeg pipeline so that the tonemapping processes are less CPU/GPU/memory bandwidth intensive as well.
It’s an impressive system for sure, but I will happily pay a price premium for the Pi ecosystem where I am unlikely to be the very first person trying to deploy any particular package.
It's an ordinary x86 board that can run any x86 Linux distribution. You are vastly more likely to be the first person trying to deploy any particular package on a Pi than an x86 computer.
So with this board, I can just plug in a LibreELEC image and start watching digital media on my TV from my sofa, while controlling it with my regular TV remote via CEC?
That's the kind of thing that the Raspberry Pi's combination of consistency and popularity does provide.
No, but you can run multiple docker containers on it for years without the fear of your SD card failing from excess writes as it uses standard M.2 SSDs
If a CEC compatible media player is what you need, then the Pi is a good choice. Or if you need GPIO.
Otherwise they're horribly overpriced compared to N97/N100 minipcs, with the exception of the Zeros.
I haven't had any issues with SD cards failing in Raspberry Pis.
But I've only got about 14 machine-years of using them so far. On a long-enough timeline, I'm sure that failure is inevitable.
The various Android phones I've had over the past decade and a half, however: Now those have killed some SD cards.
YMMV.
(BRB. I think imma dump RetroPie onto an SD card, plug in a PS3 controller, and do some trouble-free some vintage gaming on the Raspberry Pi that was doing media center duties with CEC a minute ago. If I have time, and it seems fun-enough, I'll get the controller working with Bluetooth.
And then I think I'll plug the LibreELEC card back in and crash out on the couch watching a movie.)
> the 40-pin GPIO header handled through a Raspberry Pi RP2040 microcontroller.
This part would be a reason why: very low GPIO throughput since you're limited to what can be shoved through serial and it all has to be must be manually processed by the extremely slow RP2040. The main upside with something like a Pi 4 is that it has zero overhead access to it via kernel api, so you can pipe hilarious amounts of data through its pins or address i2c devices directly from python with existing libraries for almost every sensor, etc.
The Pi 5 using it through a PCIe southbridge is quite a lot worse, especially since it broke all compatibility and it still can't even drive even a simple WS2812 led strip yet a year after release. They've thrown away the last 10 years of community support and development since the layout was standardized with the Pi 2.
The GPIO on RPi is not very useful for precision work and you’re limited to using SPI (usually to talk to an auxiliary microcontroller). The GPIO on RP2040 is so good that you can use it as 24 channel 100Msps logic analyzer in a pinch.
Smaller CPU aside, how do its GPIOs compare performance-wise to the older PRU contained in TI Sitara CPUs used in BeagleBoards? Many complained about the small number of channels on TI processors (only 4 if memory serves) therefore I wonder if it could be considered a successor where very fast digital I/O is needed but power and memory to run Linux is not because it's being hosted and run on a bigger nearby CPU.
The X4's biggest problem is the Pi form factor, which yeah, is Radxa's mistake, but also a good reason not to buy a Pi 5, cause you'll also need a heatsink, which'll then interfere with your GPIO access, fan header, and so on. If I have a small eInk display HAT, I'd much prefer not sandwiching the hot bits between the really hot bits and my display.
According to your link, this board doesn't have built-in PoE, but requires an "optional hat". It links to a "currently unavailable" listing on Amazon, without any price info.
However, I agree that if it costs as much as a RPI, it looks really great.
It says "PoE Support", but requires a separate HAT (the transformer alone would take up too much board space to fit on the Pi-sized footprint). It has a 4-pin PoE header just like the Pi.
I would imagine it would need at least PoE+ with a 30W power budget.
Not the parent but in developing and testing server software it's necessary to have the target architecture available because all sorts of random crap changes between architecture to the point that once things become more than trivially complex, the probability if things breaking when the CPU type is changes becomes close to 1.0.
I would hardly call hardware video decode/encode overrated. A hardware block will do the work with a fraction of the power that the CPU would take to do it, and in the case of something like a Pi 4 or Pi 5, you're now spending all of your limited CPU on video, rather than doing something useful.
Maybe not a problem when you're plugged into the wall, but for those of us using Raspberry Pi's (and other computers) in mobile situations (laptops, robots, drones, etc.), the power consumption is a huge concern.
My main usage of the Pi 4's video encoder was to record the video from the camera "for free" while the CPU was nearly maxed out doing stuff in OpenCV.
You misunderstood what I said. Transcode is not necessary. Pretty much every device that comes in the Happy McMeal box comes with HEVC and h264 decoding capabilities. So you just stream the data directly to the device which does the decoding with little energy as you mentioned. Transcoding is unnecessary.
Some folks have media libraries that can be used outside of their homes, and sometimes they use them with very limited bandwidth.
With transcoding, I can keep my high-bitrate 4k scan of Steamboat Willy at home, and stream the content to my pocket computer with the shitty free wifi on a public bus if I feel like it.
Another option if you need single threaded performance along with decent graphics is used corporate mini PCs with Ryzen 5 Pro 2400G CPU from HP, Dell and Lenovo. These are just starting to hit four years since first volume availability this Fall so corporate fleets are increasingly being recycled on eBay. They're already down to around $100 delivered with SSD and Win10 Pro installed and prices should fall even further in coming months as supply increases.
These are terrific for retro emulation as they can easily handle emulating everything up to and including PS2 and Gamecube at 60+ fps (with upscaling too). Toss in an old 2TB HD and it'll hold a library with many of the greatest games ever made that'll take a lifetime to play. Also, don't forget to check out the huge variety of community mods, HD texture packs, cheats and fan translations now available. Around our house we're pretty much opting out of the DLC, multi-service-login-required, 100GB-bloated, DRM hellscape that much of current gaming is devolving into and actually enjoying it a lot more.
The pi 2 and 3 were very well priced for what they did. The 4 got hit by covid and is still selling at bonkers prices - more than the 5 in some places :D
The 5 is a powerful machine, but also very expensive. Add a case, M.2 hat, power supply and you're hitting 150€ easily.
Unless you need something very specific, like HDMI CEC, GPIO or the ability to swap to a new board in a bigger industrial device - the 5 is not a good choice.
Any N97/N100 chinesium minipc or a used corporate machine like the Lenovo M700 will outperform the 5 in every metric along with having stadard M.2 + SATA + SoDIMM slots for upgradeability.
I would like to add another comment on the matter of idle power draw, that is turning out to be somewhat disappointing on the platforms of this generation (a).
The TinyMiniMicro PCs from 8/9th-gen Intel are impressive as they can get below 3 W at the wall. Gemini Lake (Refresh) thin clients can also easily get under 4 W. I wonder whether these performances come from optimisations driven by actual market requirements (I guess it can make a difference for a company that run thousands of those?).
ARM RK3588 platforms are champions in this regard. My Orange Pi 5 Plus idles at 1.5W (less than a RPi 4, with the same power supply). However, they are not viable for people who want to run Proxmox.
It seems that almost all N100 platforms idle above 5 W. Also, there are not so many passively cooled options from reputable manufacturers, while actively cooled boxes are not so silent (while TinyMiniMicro PCs are super quiet under most usage scenarios).
(a) With EU energy prices, a 5 W difference translates to 20-35 EUR per year on the electricity bill.
14c here in Spain yeah. Some countries like the Netherlands are a lot higher but they painted themselves in a corner with gas dependence. At the same time Russian imports became politically unviable their own gas fields became untenable due to earthquakes. Prices went up to 50-100c per kWh and I believe still hover around the 40c level there.
Most of the electricity there isn't generated by gas but they also stimulated electric heat pump heating heavily leading to a huge pressure on the grid.
I'm paying 6.9 cent per kWh (yes, 0.069€, I've counted zeroes right) in the Netherlands right now. It is sunny day with almost no wind. It will be more expensive at night (my provider adjusts prices each 30 minutes). But highest price in last 7 days was 17с/kWh around ~22:00 at Sunday :) Typical not-midday price is about 7-8 cents.
I don't have market pricing here in Spain, I have fixed per kWh pricing only. I prefer this because I hate the min-maxing and figuring out cheap hours etc. I just want to use my energy when it suits my life, I don't want to rearrange my life around it. But the people that are totally into that love it because it allows them to save cost. I really don't care.
I think you have to get market pricing if you have solar panels but most people don't bother especially because that market pricing makes them totally cost-ineffective. Also, I live in an apartment anyway where I can't even have them if I wanted to.
Yep. But it is still cheaper around the year than "old-style" providers with fixed or per-month prices. I could pay 100 euro/month for electricity (for everything but hot water and warming) and gas (for hot water and warming of house, double-contour gas boiler) in summer and 250/month in winter, but typical "fixed-price" contract is 200-210/month for my house each month no matter what.
And fixed part of these 100-250 Euro (like grid maintenance, which doesn't depend on consumption) is rather large, something about 75 Euro/month, so in summer I'm paying almost only for grid :)
Hmm yeah but my price is constant 24/7. I don't have to deal with that flexible stuff :) I don't want to change my life according to the energy market.
Did you undervolt it? Or you just got some blessed silicon.
7700 desktop was drawing 35W idle for me so I would love to get anything close to this magic.
It's a mobile CPU 7040 series and you can't undervolt it. But you can limit TDP using a tool like ryzenadj.
I set it to 20W on battery and the full 51W (or what it was) while connected to AC.
Not sure about lottery because this is my only Ryzen PC, I think a 7840U might run a little lower though, if I'm not mistaken about the binning process.
Take note this is a real idling wattage when it's not doing much at all. At the time when I tested this yesterday, I only had a terminal and browser open (Linux), and turned the screen off via dpms.
This is blessed silicon for sure mine ryzen always (3000 series 7000 series) have very high power consumption in idle right now I’m mostly on M1 ultra and are around 7-10W idle compared to 45W on 3900x and 35-40 on 7700. (I have also intel server with 12700k that idle around 15W but it is undervolted a bit)
The N100 seems to be everywhere - presumably it’s some kind of major leap forward in perf/watt or perf/$ versus, eg, J4125? For those in the know, is an N100 mini-pc currently the best place to start for a kid’s “my first Linux PC”?
The main advantage of the Alder Lake N CPUs, like N100, N97 etc. is that they are very cheap, cheaper than the Arm CPUs of comparable speed, so they have made the latter not competitive in most cases, because they are both slower and more expensive and only their somewhat lower idle power consumption can make them preferable in certain cases (a properly configured Alder Lake N computer should have an idle power under 5 W, but going much lower than that is unlikely).
A few months ago Intel has launched a refresh of Alder Lake N, with the code name Amston Lake, which are branded as belonging to the Atom x7000 series and Radxa said that there will be future variants of Radxa X4 that will use Amston Lake.
An alternative to Radxa X4 for the cases when its minimum size and its minimum price are not essential is Odroid H4 ($100 without memory) or Odroid H4+ ($140 without memory). That has the Nano-ITX size (5" by 5") and a big heatsink that can work even passively, fanless.
The Alder Lake N/Amston Lake CPUs have the advantage of peripherals with higher speed than any competing Arm CPUs.
Radxa X4 has 3 independent 10 Gb/s USB ports plus 2.5 Gb/s Ethernet plus 32 Gb/s M.2 PCIe socket. The best competing Arm-based computers, with RK3588 or with MediaTek Genio 1200 have at most 3 independent 5 Gb/s USB ports. All the Arm-based boards that appear to have many USB ports incorporate an USB hub, so those ports are not independent and their aggregate throughput is limited to 5 Gb/s.
Another great advantage is that the Intel GPU is not only much faster than the Arm GPUs, but it has public documentation and it has much better software support.
Also, x86 is much better supported than ARM by OS and software.
A N100 inside a fanless case is a great setup. I hope some of these low-power Intel CPUs start popping up in laptops, where competition for Apple M is truly welcome.
> The main advantage of the Alder Lake N CPUs, like N100, N97 etc. is that they are very cheap, cheaper than the Arm CPUs of comparable speed...
BTW, this is the first time I've noticed Intel could compete with ARM on performance + energy efficiency + price. Has someone already investigated how this happened? Do ARM royalties went so high up or Intel does price dumping?
I've been really impressed with the N100 Mini-PCs. I have a couple running Proxmox and then a few different VMs and containers to do light server stuff. For $180 the Beelink gives you 16GB of RAM, 512GB of SSD, plenty of video output. Just a fantastic performer.
Given the "good enough" performance and low power credentials I am really surprised that N100 based laptop is not commonplace. Make a 1440p QHD laptop with 32GB RAM it will be selling like hot cakes. The advertised max RAM is 16GB RAM but some say it can takes more up to 64GB.
We tried that dance with the netbooks of yore, and people rapidly found that their "good enough" performance was not, in fact, good enough for a lot of interactive use, particularly given that you can't burst power forever on laptop batteries.
I'd love to see a rehash of that idea with the performance actually being "good enough", but I think that's probably why vendors have been burned and don't want to explore that form factor space again soon.
I strongly doubt that. Otherwise premium (by Apple standards at least) or midrange specs laptop but with a slow CPU wouldn't be that popular.
It makes sense in a very cheap (< $300) device but add 32 GB of RAM, a decent screen and chassis and it would be competing with significantly more expensive laptops with a proper CPUs. Seems like an extremely niche product.
Personally if they are selling around USD500 - USD600 I will be buying in a heartbeat and I think most people will be buying as well especially in developing countries.
The typical specs of the price range mentioned above are FHD and 8GB RAM, both features are a few times worst laptop compared to QHD dislay and 32GB RAM.
I’ve got likely the exact same box as you, and I’ve done the same thing. Mine runs proxmox with a variety of services.
Additionally I managed to get gpu passthrough working with the integrated graphics. My box is physically located below a TV- so I’ve got one VM running Bazzite for gaming, outputting over the HDMI. It’s perfect for streaming games from my gaming pc to the couch. Bluetooth and USB is also passed through to the VM without issue for controllers.
Slightly OT but Proxmox is something else I see is suddenly everywhere - is it only for a container-based homelab setup, or do people also run a full desktop distro like Debian or Fedora as one of the VMs? Since GPU passthrough is supported, seems like a nice neat way to have your cake and eat it too. In which case I might splurge on an N305 with a bit more RAM and migrate Plex and other containers (currently on an always-on M1 which is supposed to be my workstation), plus the Tailscale exit node (currently on an RPi4) to the same machine the kid uses as his DE.
It's not a container-only distro, that's secondary, it's more like VMware Esxi, does VMs, too, but with extra bits (like ceph). It's good stuff, try it.
You asked about mini-PCs, which I don't have, however just for comparison I can say I have ODROID-XU4 + cubietruck here running and only drawing about 2W total between both of them.
I wouldn't necessarily recommend those (especially the cubietruck) these days, but there are lots of Single Board Computers (SBC) out there besides RPi.[0]
Those things are insane value for the money. First breakout product intel has had in a long time, honestly.
The downsides are that it's single-channel memory, which I'm not overly thrilled about, personally. Although I understand it's an important element of keeping the cost down, both in terms of die cost and also the systems integration around it. One channel = 1 socket... or soldered memory with half the modules.
That said, the overall package speaks for itself. It's fine. Which is the standard to meet in this class of product. It costs less than $100 for a loaded rig that you plug in and run windows on, what are you really expecting here?
And when you drive down the cost that far, a whole bunch of new use-cases open up. This is what the social implication of moore's law has always been - that eventually your fridge will have a computer with hundreds of MIPS of performance, because it's so cheap why not?
And that was the argument for the RPi initially. Other than the "they'll use it to run computer labs!" which was silly and never materialized either.
Been a big fan of Gemini Lake Plus for a lot of years too. It’s cheap and good enough (and with the right modules you can stick 16gb or more (tested up to 32gb) of memory in it, which makes it totally acceptable for light desktop use. Same thing. It’s good enough.
In comparison with a Pi 5, the slowest of the above, which is the X2L/Palmshell kicks the butt of the Pi 5 six ways to Wednesday. Just so much better, it's unreal. Web-page resizing works immediately, video plays well, it just feels like a real PC, even though it's a bit slower than my EliteBook, which is still pretty handy.
Not got an X4 yet, but from the link above, it should be about 50% faster than the EliteBook, which in turn is 50% faster than the X2L / Palmshell.
I tried one of those N100 mini pcs and went back to the J4125 purely for the thermals. My needs are not great, and the fans on the mini pc were quite loud quite often. The J4125 hums along fanless with a small heat sink no problem.
It’s a bit tricky to do that, and hardly justifies the effort. I have a J-series, an N100 and an N5105, and the biggest bottlenecks are always about I/O. The N100 makes a great Proxmox server solely due to the faster cores.
Different workloads I guess but context-switching/CPU (and associated increase of memory bandwidth) is/was holding my N5xx/N6xxx back from properly utilizing I/O.
But this was in the context of someone who has an N100 but prefers running a J-series over due to lower power consumption. It seems like it should be worth the effort?
Odroid H4+ ($140 for the SBC, $10 for a case; extra cost for memories) is an example of a cheap passively cooled version. It can also use a fan, but the fan is optional and not required.
There is an even cheaper version Odroid H4, at $100, which has a single 2.5 Gb/s port instead of two and it lacks SATA ports for HDDs (the fewer peripheral controllers also lead to a lower idle power consumption).
Noticed the same thing and I hope we see the N305 from the same generation take over or more vendor offering both options. Considering the rest of the platform package, it can really benefit from 8 cores instead of just 4.
The N100 can be a fair step up compared to Rpi5 but even RK3588 is already 8 cores. Would be a shame if many of the current generation of exciting hackable x86 mini-platforms lock in at the N100 as it will feel obsolete years earlier the the N305.
I run/ran stuff on both, as well as various ARM SBCs and previous generations like J4125/N5XXX. Considering the core-count, RK3588 is still a better pick for many use-cases unless single-thread performance is that important. Benchmark comparison: https://bret.dk/intel-n100-a-challenge-to-arm/
It really matters very little that RK3588 has 8 cores, because the Cortex-A55 cores are very weak.
The reason why RK3588 may beat N100 in multi-threaded benchmarks is that the latter has a very low base clock frequency, due to the higher power consumption when all the cores are active. N100 needs a better cooling than RK3588 in order to reach its maximum possible multi-threaded speed. With good enough cooling that will prevent the drop in clock frequency, N100 will beat easily RK3588 in any benchmark, because the Intel Gracemont cores have a speed similar to the Arm Cortex-A78 cores and much greater than the Cortex-A76 cores of RK3588.
It should be noted that Radxa X4, due to the constraints of the credit card size, has only a 32-bit DRAM interface. So its -4800 memory is equivalent with a -2400 memory of the bigger N100 boards, which use a 64-bit DRAM interface. The good RK3588 boards also use a 64-bit DRAM interface (with a somewhat lower speed, e.g. -4267), so they may win in benchmarks that are limited by memory bandwidth.
However in the applications in which I am interested for such a small board like Radxa X4 (like a network router/firewall or a controller for some custom hardware) the performance is almost always limited by the speed of the peripheral interfaces, so Radxa X4 will be better than any existing Arm-based SBC.
The RK3588 doesn’t beat the N100 in anything but throughput - I have been benchmarking a number of SBCs lately to consult on building custom industrial edge devices and in terms of bang for the buck and PCI lane availability (as in actually accessible lanes) the Rockchip designs work out a little cheaper. I have one of those 4xNVMe NAS boards to test next, and expect it to be pretty snappy for the same wattage-but it all depends on the use case.
(I also have an X4, am waiting for an SSD for it for proper testing)
While I haven't benchmarked, my reasoning is that for parallel workloads (especially VMs but also containers), penalty from context-switching should be vastly reduced by even those slower cores.
Good luck getting timely mainline Linux support if ever. Rockchip usually provides an ancient patched kernel without much willingness to upstream their changes. They usually leave things for the community to sort out. On the other hand, x86 has the best mainline support which you can count for awfully long time to be around.
I recently bought an n100 and within a matter of days got buyer's remorse and impulse-purchased an n305 to go right beside it, which is currently sitting with a wildly overpriced 48 GB stick installed and 2TB SN850X, it's an absolute joy perfwise and the absence of heat it generates.
The only thing I'd reserve judgement on is the tendency to throttle. I haven't got far enough to characterize it, but it's not clear how much value those extra cores will add over the n100 with TDP settings tweaked down in the BIOS, and if leaving the n305 to run at max TDP, heat/noise/cost/temperature-related instability may start to become an issue, especially when packing other hot components like a decent SSD into the tiny cases they come in.
Any old desktop from a garage sale (if you can find one). People are practically giving them away these days, everyone wants tablets/phones now. We scored a quite decent one, complete with monitor, mouse, keyboard, etc. for like $25 (IIRC)!
In fairness, that was a few years ago, but I am sure there are deals to be had if you look around.
I would not recommend ARM/SBC for "first GNU/Linux computer" necessarily, you will likely have an easier time on x86, as much as I hate to say it.
I got my kid a raspi 400 a few years ago. She’s enjoyed it tremendously, but I don’t know how it holds up to more recent alternatives. She’s still using it today, though.
>This means that I was able to grab the Windows 11 ISO and install it. I was able to go to Debian’s homepage, grab the latest ISO and just, well, install it?
Still the biggest advantage x86 has over ARM in my opinion. I hope one day common ARM processors become as flexible.
Either way its amazing to see x86 processors so cheap, low power, and still rather powerful. I hope to see more devices using these N100's
Ampere is the main company with that kind of setup (with full EDK II BIOS and UEFI), but they still target server / HEDT, not desktop, and certainly not the SBC/hobby space.
I asked Raspberry Pi about the likelihood of seeing UEFI support this year and it didn't sound like it was a priority, at least not for the Pi 5 generation. Would still love to see it, as it would let distro maintainers focus less on a few "premiere" Arm boards and be able to just have one distro for aarch64.
I recently got a Raspberry Pi 5 and regret not getting one of these N100s instead for a small home server. A lot of the killer features of the Pi like using a traditional phone charger, being passively cooled, and being "cheap" are no longer there if you want a stable system. On the other hand, the "low power" from ARM systems, it's not that big of a lip compared to a traditional X64 mini-PC CPUs.
Yep. I was in the same spot. The N100 is also cheaper if you factor in the adapter, case, nvme etc. I had a N100 for 100 bucks. Much better than a pi for a small home assistant server IMO. Raspberries are no longer cheap nor especially good value for money.
Things were really different at the time of their introduction. There wasn't really anything similar except the Marvell based sheevaplug.
The only thing I'd use a pi for now is something that really needs GPIO as well as significant processing (eg more than an esp32 offers).
I think that's the most interesting part of this particular board - the built in RP2040 and 40 pin header for Pi hats. With the Pi 5 splitting off GPIO to an external chip anyways there really isn't much advantage to having a "real" Pi for GPIO/hats vs this - some stuff around firmware/software image integration for really complicated hats mostly.
I hope Intel continue updating/improving their low power offerings like N100. It has been a great little CPU so far. ARM ecosystem seriously lags behind x86 in terms of standardization. Hunting down ISOs for devices other than popular ARM boards gonna be impossible once the vendor ends supporting their heavily patched Kernel. Seems like that the ARMified future is not gonna be the rosy dream as vendors want us to believe.
You can use aarch64 ISOs on some ARM64 computers through UEFI, RPi3+ included. Need to install firmware though. Similar if you use Tow-boot on some Pine64 devices and there is a project for recent Rockchip boards.
I have 2 Rockpro64s which I boot and installed the default ISO of Fedora Server from an ISO file on USB.
$60! That seems like a ridiculous bargain. Evidently I’m not the only one who thinks that, because they are sold out. Not in love with the reported temps, but I would probably be willing to throttle the CPU significantly.
Does anyone have experience of putting a huge GPU on something like this and using it for inference? You'd be limited by data feeding over the NVME port, but otherwise you won't be bottlenecked right? Seems like a light weight and cute way to limit non-inference power/weight without having to pay the price of a Jetson board.
You just don’t have the bandwidth to do that. Even if you use the M.2 slot you’ll be significantly bottlenecked and would be better off using something else - even an AMD iGPU will work better (https://taoofmac.com/space/blog/2024/04/13/2100)
What’s the bottleneck? Once I’ve got the model and data onto the GPU my only cost is launching CUDA kernels right?
Not sure if that blog post is relevant, but even if it is it shows a 3060 gets //way// faster throughout than the igpu it is testing. I suppose I can test this myself by plugging my 3070 into the NVME on my desktop.
It’s a mobile platform, so saving 100 w on the CPU would make a difference. That’s the answer I tell myself, the real answer is because it would look hilarious :D
One of these with Emulation Station should make a fantastic retro gaming machine! I'm using an Rpi 4 as one at present, but for sure it would be better with much more CPU power.
As the other commenter noted, I've recently been migrating from retropie to batocera and its night and day for the most part. I've had more quirkiness on batocera due to it being less of a standard platform, but the performance itself is not even close. Full upscaling on retro consoles, and easily handles nintendo switch.
Im validating it on a desktop pc at the moment, but have a minipc in the mail that should handle it fine (minisforum um790 pro).
No links for pre-loaded if you mean roms, but batocera is a much more pre-loaded experience if you mean emulators and config. Its a much more rigid experience than retropie. You can get it at the homepage: https://batocera.org/
The nice thing is that it is designed to be run from a usb-stick, so you can easily pop it in and validate the use case before going full-install.
SATA is awkward because the drives need a +12V rail, and this board probably doesn't have one, so they'd have to cram another power supply in along with the SATA controller or leave the user to bodge together an external power supply solution. If you don't mind doing the latter you could put a SATA controller in the M.2 slot:
The board is powered via USB-C PD at 12V, so it might as well have a 12V rail.
But with a SATA connector they probably would have to account for power hungry devices with spinning disks, which feels a bit antithetic to having a power efficient processor.
I think they just had to make some compromises for price and form factor. There are a lot of other boards out there based on the N100 or its close friends. You might like the ODROID-H4+. https://www.hardkernel.com/shop/odroid-h4-plus/
The advantage of the Radxa X4 is that you can buy it instead of a Raspberry Pi: similar price, similar form factor, similar GPIO options via the on-board RP2040.
Can't wait to get this & have TrueNAS+Jellyfin up & running on it, certainly beats having to buy so many addons for the Pi & still not have a super reliable build.
Absolutely could not help myself, preordered one. Same speed as my always-on i5-6500t mini-PC, and has a RPi0 on it? Neato. Coherent boot system & mainline kernels? Oh heck yes so much better than this amateur hour most cores make you suffer.
Had kind of been planning to get whatever comes next, at a bigger tdp (more N305 class), but Radxa really nailed the price here, while being very fully featured.
I'm super glad to see an x86 hardware WITH PoE as an option. That seems to be a niche almost exclusively found on arm boards, but OS selection kinda sucks for Arm SoCs (at least the OS I care about, NixOS, doesn't have very good out of the box support for raspberry pi 5 yet).
Genuine question, but what's the use case for this? I do a ton of embedded compute, but it all requires extensive GIPO/PWM, etc. Then I do a lot of desktop compute, but it all requires a GPU and a fair bit of horsepower.
What's the use case for a SBC made to be embedded in a project box, but without extensive IO?
Emphasis on the low power. It’s just a computer, but one that is cheap and yet still capable of doing work.
Host a blog. Home automation. Dedicated SNES emulator. Personal Minecraft server. DIY NAS controller. Maybe good enough to act as a TV media player. Anywhere you might want a bit of always on compute without paying a ton.
Intel N97 seems to support a so called In-Band ECC with non-ECC UDIMMs, thus there is a slight performance hit (both memory bandwidth and processing resources). I'm curious whether N100 does support it too, though considering a huge difference in a recommended price, it could support, but it would be artificially forced disabled.
This does have provisions for GPIO, there's an RP2040 microcontroller integrated onto the board which is wired to the N100s USB bus on one side and the exposed pin headers on the other. It's set up such that you can flash your own code onto the RP2040 so it should be quite flexible, you could just use it as a simple USB to GPIO bridge, or you could implement low level application logic on the RP2040 itself for stronger realtime guarantees.
I wonder how good the software/driver support is? That's a neat feature to have on an Intel-based SBC, but not if getting software to talk to it is next to impossible.
What do you mean? I looked up the RP2040 datasheet[0], and it looks like it can be used as a plain USB device, and can be booted from USB. If that's how it's connected on this SBC, I'd imagine software support would be quite good since Linux probably has a built-in USB driver for it.
I'm talking about how the RP2040 and the N100 are interconnected. With an UART port and also an USB port. Someone mentioned the USB port for software updates but maybe both UART and USB port can be used for communication between them.
Linux doesn't have any driver for the "RP2040" in the abstract - you have to design the USB device you want (which you could implement on any other USB capable microcontroller as well usually.)
You can use the standard USB CDC classes with the RP2040 to obtain Linux support without writing your own driver - but you still have to build this.
You can only access the USB boot ROM after programming if you are able to toggle the boot selection pins with the host CPU's GPIO - no idea if that's connected here.
Raspberry Pi sized computers are very popular as second systems for experimenting and as home servers on a budget. For many people, it’s an easy and cheap way to get a Linux server running in their home, apartment, or dorm, without spending a lot of money or having a noisy, hot, large old server box.
They do okay for a lot of tasks. This board has a decent GPU for the power envelope. They can’t keep up with a $1000 workstation or server build, but it’s a lot of horsepower for under $100 and around a dozen watts.
There are 2 kinds of uses for this. Those that require a very small computer and those that require a very cheap computer.
You can buy 4 or 5 Radxa X4 for the money that would be used for one NUC-like small computer with an AMD CPU or with an Intel Core CPU.
Radxa X4 is much slower, but there are applications for which it is good enough, so there is no need for a more expensive computer.
For example I would use it to replace an old Intel NUC that I am using (together with several USB Ethernet interfaces) as router/firewall/DNS server/DNS proxy/e-mail server/Web server/NTP server etc.
Radxa X4 would be good enough to ensure the routing and filtering for four 2.5 Gb/s Ethernet ports, while cheaper devices like those using quadruple Cortex-A55 cores are not good enough.
For that Internet gateway, which must work 24/7 without downtime, I keep a spare unit ready to replace the main unit, if necessary. That means a double cost. With Radxa X4, the cost for two units remains very small, which is an incentive for using it instead of a more expensive SFF computer.
The alternatives at the same price, i.e. Rasberry Pi and the similar SBCs that use RK3588, have a much lower peripheral throughput aggregated over all their peripheral interfaces than Radxa X4. With external USB hubs, Radxa X4 could be extended to 8 full-speed 2.5 Gb/s ports or more, or I could use up to three 10 Gb/s Ethernet ports instead of the 2.5 Gb/s Ethernet ports.
The small size of Radxa X4 would be handy if I would use it as a portable firewall for a laptop. Unfortunately the modern laptops cannot be trusted due to the hardware backdoors that are implemented in them for remote management. So a paranoid user would want to disconnect the internal antennas and connect to wired or wireless networks through an external firewall, for which a Radxa X4, which has a published schematic, would be a good choice.
The first batch of Radxa X4 are all sold out. And X4 is under FCC compliance and Aliexpress is not allowed to list the product without FCC ID if no stock.
The main advantage of Raspberry Pi RP2040 microcontroller on Radxa X4 board is that they can now use the Raspberry Pi logo next to the Radxa X4 marketing to increase the awareness that they are "similar to Raspberry Pi". Without RP2040 they wouldn't be allowed to do that...
Went down this route and haven't had any issues with heat or having to replace heatsinks, add fans, etc. that others mentioned, but your mileage may vary.
Runs opnsense on proxmox, along with some other containers. It's been a great little box.
It has only 9 (old pcie 3.0) lanes though which is bit limiting. In theory you should be able to use 8 of those lanes and get something like 4x10Gbe, which would be actually really neat, but that requires that the lanes are not squandered for anything else.
Curious about the max frequencies on the GPIO pins since I believe it drives them with an RP2040. Not actually sure how the RP2040 interfaces with the Intel chip though.
I presume it would just be attached over USB internally. That would make using it with the official SDK and other tooling as easy as a Pico attached to any other computer.
A new hardware project should better use the refresh of Alder Lake N, i.e. the Amston Lake CPUs, which are sold as the Atom x7000 series.
Alder Lake N and Amston Lake are pin compatible, so the CPU model does not influence the schematic and PCB design.
The development material is given by Intel under NDA, so you must contact Intel sales for that.
Nevertheless, you can see the equivalent of a reference design by downloading the schematic of Radxa X4 and its PCB assembly drawings. Radxa is among the nice SBC vendors who provide good documentation for their products. All their products come with their complete schematics, as they should.
For those who do not want or cannot enter in an NDA relationship with Intel, it is possible to design a product that uses a SOM instead of a CPU package. This greatly simplifies the design and the manufacturing, but SOMs are intended for industrial applications, so they are expensive.
For example, I have looked now at DigiKey and I have seen a Advantech SOM with N97 (which is better than N100) at $272, with 4 GB DRAM, which is more than 4 times more expensive than a Radxa X4 SBC with 4 GB DRAM @ $60.
The cheapest way to make a hardware design with Intel CPUs remains to integrate a small SBC like Radxa X4 into a bigger box, together with a custom PCB.
Arm CPUs are not better. To make a hardware design with MediaTek Genio 1200 (quadruple Cortex-A78) or Rockwell RK3588 (quadruple Cortex-A76), you still need to get under NDA the required documentation. Only for microcontrollers like those made by Renesas, Infineon, ST, NXP, Microchip etc. you may find the complete documentation on-line.
For RK3588 there have been published many schematics of various SBC's, so for some less important project one could design a PCB based on the public RK3588 datasheet and on the available schematic examples, but for a serious project one would want the real Rockchip documentation.
In the same way, for an amateur project one could make an N100 design starting from the published Radxa X4 schematic, even without the Intel documentation. However, the Intel CPUs like Alder Lake N or Amston Lake are normally sold in bulk, e.g. 1000 pieces or at least a few hundred. It may be difficult to find a distributor who would sell smaller quantities.
> Arm CPUs are not better. To make a hardware design with MediaTek Genio 1200 (quadruple Cortex-A78) or Rockwell RK3588 (quadruple Cortex-A76), you still need to get under NDA the required documentation. Only for microcontrollers like those made by Renesas, Infineon, ST, NXP, Microchip etc. you may find the complete documentation on-line.
Perhaps less interesting, but there are a handful of Cortex-A SoCs from TI [0] and NXP [1, 2], at least, that have non-NDA datasheets and reference manuals.
True, but all those are really obsolete Cortex-A cores.
The Cortex-M cores used in microcontrollers have a much slower evolution, but the Cortex-A cores are updated every year and the differences between the more recent cores and the older cores are very big.
Cortex-A78, like in MediaTek Genio 1200 is the 2021 Cortex-A core.
Cortex-A76, like in RK3588 and Raspberry Pi 5 is the 2019 Cortex-A core, already 5 years old.
All older Cortex-A cores are hopelessly obsolete.
There still are some borderline acceptable SBCs that use Cortex-A73, the 2017 Cortex-A core, i.e. 7 years old.
The NXP and TI products use much weaker Cortex-A cores than even that.
The only NXP products with Cortex-A cores that can be acceptable for certain purposes are those with Cortex-A55 cores (2018). These at least support the Armv8.2-A ISA, which corrects some important defects in the original Armv8-A instruction set.
Nevertheless, the Cortex-A55 cores are very small and slow, so using a CPU with them is justified only if that results in a much smaller or much cheaper SBC than when using Cortex-A7x cores. That means that any acceptable SBC that uses Cortex-A55 cores must be small and fanless, e.g. a fanless credit-card-sized SBC that must be cheap too, e.g. not exceeding $50 unless it contains a lot of memory and interfaces that could justify a price slightly higher than that.
Many of the solutions using TI or NXP Cortex-A based devices fail to satisfy the power consumption, size and price values that are expected for such cores (frequently due to their use of ancient manufacturing processes, which may make a Cortex-A55 consume as much power as a Cortex-A7x made with a newer CMOS process).
NXP makes microcontrollers with 1 GHz Cortex-M7 cores, which are good enough for many of the applications for which Raspberry Pi SBCs have been used. For the remaining applications, which need fast interfaces, like for SSDs, USB 3 or multiple Ethernet interfaces, it is better to go directly to a SBC like Radxa X4, so there are only very few applications where a device with obsolete Cortex-A cores can be preferred to a either a cheaper microcontroller or a faster Intel Atom CPU.
> The development material is given by Intel under NDA, so you must contact Intel sales for that.
Any idea on MOQ? Do they still require supporting chipset purchases?
Cost, pinout and voltage inflexibilities usually preclude my ability to use SOMs.
The RK3588 documentation is all available without NDA, as are a handful of Rockchip's reference Altium designs. I've had no issues with the mainline RK3566/RK3568 - love these chips.
Searching now, I see that Atom x7425E (which is almost equivalent with N100, but it has a few extra features enabled, like in-band ECC, and it may happen to have a higher power consumption at default settings, because it is not tested @ 6 W, like N100; but for any Intel chip you can easily configure the desired power limits) can be bought as single pieces at Avnet for $79 (drops to $71 for 100 pieces) and at Mouser for EUR 79 (drops to EUR 67 for 100 pieces).
The Atom CPUs, like N100 and the other Alder Lake N/Amston Lake, do not use chipsets.
You can look at the Radxa X4 schematic to see what may be needed for power management and clock generation, besides the CPU package.
If you only need small quantities, the Atom branded Alder Lake N are easier to find than those branded as N100/N97/N200/i3-N305, which are normally sold in large quantities to the big manufacturers of computers. You can see the available models in Intel Ark, at the Atom x7000 series (Atom x72xx are 2-core, Atom x74xx are 4-core, Atom x78xx are 8-core).
RK3568, which is a quadruple Cortex-A55, is in a completely different class of products. Any price over $20 would be unacceptable for such products.
The chips competing with Intel are only RK3588 and MediaTek Genio 1200.
Those cannot be cheaper than Intel, because the SBCs made with them have prices from the same price as the Intel SBCs to much higher prices than the Intel SBCs.
Radxa, which buys N100 in bulk, must pay at most $50 for one (the Intel list price is $55), but probably much less, perhaps as low as $30, but when you want only a few pieces you cannot buy at that price.
LCSC has the RK3588S at 60 USD, so the high Intel prices on Mouser might just be artefacts of general parts suppliers selling specialised high-cost parts with high markup.
I thought the n100 would be a joke compared to an an HP G6 mini with i5-9xxx (small book sized computer), but seems they might be competitive with n100 drawing 1/6 the energy?
Is that right? Can n100 do fast SHA256 and h265 / x265 transcoding across multiple containers?
If you want a basic idea of an N100, it's essentially a single Alder Lake E-core cluster (4 E-cores) with one of the low tier iGPUs and a trimmed down memory controller.
The presence of the iGPU means it /should/ handle video encoding surprisingly well for what it is.
Quick Sync on Alder Lake has hardware encoder support for HEVC 8, 10, and 12bit and decoder support for AV1.
Anecdotally, my busy Plex server (Alder Lake but not an N100) routinely ends up doing 3-4 concurrent transcodes, often involving high bitrate 4K HDR content. The Quick Sync encoders take almost all of the load off the CPU and never have trouble keeping up.
That said, I haven't had great experiences driving a display with 4K+ HDR content on an Alder Lake iGPU. At 60hz it's not too bad, but it can get stuttery past 100hz particularly in high contrast scenes with lots of movement.
You are not wrong. However x86 is dead to me for having a BIOS level backdoor built in that you can't get rid of (probably; yes I know about ME Cleaner). Not to mention Intel sat on their laurels for years. ARM are far from perfect, but maybe some competition gets things progressing again.
The n100 is at least “not awful” at 3d graphics in my experience. It doesn’t compare to a discrete GPU of course, but it can handle basic games. Mine has been able to run dolphin at the consoles’ native resolution. Although I think I had occasional stutters in some circumstances.
I have an n100 mini pc hooked to a TV and I play couch coop games on it occasionally. It can do reasonably well tbh, but for simplicity I do end up streaming most games from a more powerful machine. I’ve heard performance is better in windows but I’m running mine slightly handicapped- as a proxmox VM with GPU pass through, and not a lot of ram. This same box also runs all my home automation and a few other services.
A discrete gpu is like AMD/Nvidia cards. They are very capable for graphics, and essentially required for any higher end gaming.
This machine would just have integrated graphics. Which aren’t the best but they are okay-ish. Better than I expected to be honest.
I think AMD probably has better integrated graphics? I’ve heard great things about what the Steam deck is capable of for instance. But the N100’s price point is very hard to beat.
The reason why integrated graphics usually sucks is because the people putting the GPU on the chip expect you to put a dedicated card in anyway.
An integrated GPU has to compete with the CPU cores for power budget, memory access[0], and die area. Intel doesn't want to risk half the die on a GPU nobody will use when they could sell more cores or cache instead. AMD instead manufactures separate "APU" product lines for people who want integrated graphics[1], and those have smaller core counts as a result. E.g. a Ryzen 8700G, the top of the APU stack, only has 8 cores while Ryzen 7950X3D will go up to 16. Apple's M-series chips[2] have similar core counts. The Steam Deck is even more graphics-specialized, having only 4 CPU cores in total.
Intel would be wise to copy AMD's strategy and just completely omit the GPU from their desktop chips, save for a separate integrated graphics SKU. I suspect OEMs wouldn't like that, though, because it would require putting a bunch of e-waste GPUs[3] into high-end general-purpose systems. AMD traditionally has made a lot of enthusiast and system builder sales but I wonder if that will change in the future.
The kinds of applications N100 is built for is going to be more graphics-heavy. I expect these chips to find their way into a lot of weird industrial shit, e.g. digital signage controllers, touchscreen kiosks, etc. where CPU performance doesn't matter but you need a display output and some graphics acceleration. The CPU is absolutely going to be bottom of the barrel anyway, so it's not hogging as many resources from the GPU.
[0] CPUs want low memory latency since they run latency-sensitive code. GPUs want more memory channels and bandwidth since they run everything in extreme levels of batch processing and turn-around times don't matter.
[1] AM5 Ryzen CPUs do have display outputs now, but AFAIK there's no hardware acceleration on it, you need a dedicated GPU anyway.
[2] Excluding Pro/Max/Ultra. They are unique in that they are very large workstation chips with integrated graphics. They basically don't have any competition, everyone else just wires separate dies together at this scale, even in mobile workstations where this would make sense to do. Apple is so married to the integrated everything mentality that they even configured their MMUs and PCIe controllers to forbid external GPU access.
[3] Anything in the $40-100 range. If you just need a display output you can buy cheaper and if you want to play games anything even slightly more expensive is a dramatically better value.
Edit: another source with better image of the form factor. It might actually fit a raspberry case.
https://www.cnx-software.com/2024/07/19/radxa-x4-low-cost-cr...