AMD Publishes XDNA Linux Driver: Support for Ryzen AI on Linux
phoronix.com
phoronix.com
As an OSS contributor, that available software (via tarball from their website) allowed me to make public Docker tooling and other projects [2]. I would be less inclined to do so with proprietary binary bundles.
But it wasn't "open collaboration" until recently -- but they were great on support emails! However, new GitHub projects are popping up from them even in the last few months! [3]
And thus, I can independently and openly work on issues [4] and integrate their software with other tools [5], just yesterday.
This all echoes to the recurrent Supabase thread of the value chain of OSS in companies[6]; there's a lot of reasons to do it.
[1] https://web.archive.org/web/20180108154033/http://www.openon... [2] https://github.com/neomantra/docker-onload [3] https://github.com/Xilinx-CNS/sfptpd [4] https://github.com/Xilinx-CNS/sfptpd/issues/6 [5] https://github.com/neomantra/nomad-onload [6] https://news.ycombinator.com/item?id=39087837
Is there any strix besides the unreleased Strix Point and Strix Halo?
However the AMD CEO has confirmed recently that Strix Point (laptop Zen 5) will be released later this year (and it is currently being tested by AMD partners like Microsoft) and that it will include a NPU much faster than that of Hawk Point (Ryzen 8040 series) and Phoenix (Ryzen 7040 series).
As Strix Point is currently in testing at external partners, they must have at least preliminary versions of the software for it.
https://ryzenai.docs.amd.com/en/latest/
https://xilinx.github.io/mlir-aie/
Looks like it's another NPU/TPU. Apparently it has a similar architecture as Versal.
VLIW vector units optimized for AI (-ML suffix) with runtime reconfigurable network on a chip letting you optimize "streaming" the data between individual cores.
The way "AMD IPU" device was implemented, which embeds the system in some new AMD CPUs, the previous drivers (which were Linux/RTOS only) didn't work.
I was actually in the middle of reverse engineering how it was exposed in Phoenix (7940hs) APUs to write custom driver for Linux based on the one shipped for windows.
Which makes sense, because under Xilinx brand they have been selling accelerators and SoCs using AIE and AIE-ML cores for various use cases in embedded world, and XRT lets you program those with pretty much plain C++.
For some AI tasks, they have provided high-end wrappers for models that use ONNX (iirc) so those can be used immediately. This is not ROCm, and has essentially none of the ROCm bullshit in terms of support. I seem to recall there's some work to integrate parts of ROCm (opencl, HIP) with XRT.
Definitely going to build some kernels with this and test drive it ASAP.
I have a 7900 XT, and all my attempts at getting ROCm to work on different Linuxes seem to have failed at different points. Does anyone have a pointer to a clear explanation on how to get it to work?
Honestly, it just stinks of incompetence, or if one were otherwise inclined in beliefs, disgust for their customers.
If AMD had been hiring japanese to make this, they would have had to close up shop as they would have all had to perform seppuku.
It "works", but it's an obvious second class citizen, and the developers don't seem to take bug reports / PRs seriously.
My anecdotal evidence with an AMD GPU is there are serious problems trying to use a graphics card to run graphics and throwing a few compute tasks at it - because X11 will reliably crash and the kernel graphics driver enters some sort of corrupt state. I think that is a use case that was beyond their design considerations to start with.
There is a hint that servers are targeted because they supported linux first with ROCm, and had a very tight support matrix of what OS they expected people to use. Unfortunately, this strategy generates a lot of bad publicity and I'd guess has trouble gaining traction because people can't experiment cheaply with AMD hardware. They're progressing though and even with a dose of cynicism about their wherewithal there is a lot in the pipe that has the potential to take on CUDA. Even APUs are promising.
"Hello world" in torch? On a new install/card/config/even upgrade with a bit of fiddling I can get there (I have it more-or-less down at this point).
Actually using something in a real world use case? Hours, days, or give up. Often the third scenario where I end up going back to CUDA because I just need to get something done. Messing around with different driver versions, docker containers, ROCm versions, obscure environment variables, weird random cherry picks/patches and hacks from spelunking a bunch of GH repos/discussions/PRs/blog posts/etc. Feels like it never ends.
I talk about it quite a bit and I keep buying AMD hardware and putting time into this because I'm rooting for AMD. However, consider the CEO of Nvidia saying 30% of their cost/spend is on software. He gets it.
I'm pretty convinced at this point that AMD just doesn't have software in their DNA, with even die-hard "Team Red" Windows gaming users frequently complaining about the quality of drivers. Drivers. For gaming. On Windows. Enter >80% Nvidia market share in that market and >90% market share in ML.
They just don't get it and I'm pretty sure they never will. I expect Intel, Apple, and even new entrants in the field will be the ones to offer viable alternatives to the Nvidia/CUDA monopoly.
Not contradicting you about the other areas, although I do have some hope because things (well... mostly hardware) have already improved in the GPU department under Lisa Su.
Of course with the AI goldrush they're paying a lot of lip-service to ROCm (pumping stock price?) but a couple of examples:
1) They just added official support to ROCm for the top-end consumer card they released over a year ago. Meanwhile in CUDA land Nvidia back-ported support to CUDA 11.8 well before Ada Lovelace/Hopper even launched. Worked with everything out there day of launch. Even newest CUDA 12 and their drivers support anything with Nvidia stamped on it from the past 6-7 years. Literally everything.
2) Just check their repos, Docker Hub, etc. It's embarrassing for them. They used Python 3.10 for ROCm 5.7 because its basically the bare-minimum/standard for any broader AI, ML, etc project. Oh neat, they released ROCm 6 to support their year old ~ $1000 card! Python 3.9? Ubuntu 20.04? What the hell? Pytorch "hello world" usable but almost nothing else even runs.
The further you dig the more clownish it gets.
Source? According to this https://www.custompc.com/amd-fsr-2-vs-nvidia-dlss-2-comparis... it's only a bit worse:
"However, the results were closer than this might suggest, with both upscaling modes often doing a decent job and delivering very similar levels of image quality. It was just the final 5% where DLSS 2 was that bit better."
> And the end result? Nvidia’s DLSS 2 comprehensively beat FSR 2, coming out on top in almost every comparison and generally exhibiting less ghosting and other visual artefacts while preserving more detail.
If I were on an AMD hardware team I would be (mentally) screaming at the software side of the house daily. I can't imagine how frustrating it must be to design and deliver such capable hardware only to have the software render it practically unusable.
Debian formed a ROCm Team a while ago and we are working hard on getting this to work out of the box, as in: apt-get install and you're done.
For various reasons, our initial focus was on RDNA2 and CDNA2 and earlier, mainly because we are building a CI network in which we test all of our packages (and their reverse dependencies) on actual cards, and we had to bootstrap the network infrastructure first. CI is central to the Debian ecosystem but all our official infrastructure cannot deal with the requirements for specific hardware attached to a machine -- that was never a factor, until now.
AMD has generously supported us with hardware donations, among which are RDNA3 cards. These are already in our possession, and it's just a question of person-time until these are integrated into our CI. We are already spec'ing the systems, though.
https://www.phoronix.com/news/Radeon-RX-7900-XT-ROCm-PyTorch
(Apologies I missed your response earlier as this thread suddenly had a lot going on.)
After installing the dependencies using apt go to pytorch's getting started page [1] select Stable/Linux/pip/Python/ROCm and run the command
[0] https://rocm.docs.amd.com/projects/install-on-linux/en/lates...
As with all other Debian development, we target the 'unstable' distribution for packaging work (similar to git HEAD). Packages eventually migrate from 'unstable' to the 'testing' distribution once they pass certain checks. And when Debian is ready for a new release, 'testing' will basically be tagged as '13 (trixie)'.
Ubuntu periodically syncs their packages from Debian 'unstable', so eventually they will have they same packages that Debian has.
Either way, we realize that running bleeding-edge distributions is probably not for everyone, so we intend to provide backports for both Debian and Ubuntu. Hence why I said earlier that eventually, it shouldn't matter. With "eventually" being sometime in the next few months.
(Site note: our own PyTorch does not yet ROCm support, we're still working on a few dependencies.)
Can I clarify something. If we install a stable release of Ubuntu now, we wouldn't be able to get PyTorch to work with a 7900 XT/XTX? At least not until you've worked your magic and that's made its way to Ubuntu in a few months?
You could still use a version provided by a third party, for example directly from pytorch.org, or a containerized version provided by ROCm upstream [2].
https://www.phoronix.com/news/Radeon-RX-7900-XT-ROCm-PyTorch
I have the same card, the above works with everything I threw at it so far. Haven't even installed amdgpu-pro drivers btw, only have radv (that steam installed by default).
If pytorch, keep in mind that most of the pytorch+rocm version is self contained, and you don't need to install every roc* package on your system for it to work. This confused me in the beginning at least.
Edit: Example garbage image: http://oirase.annexia.org/tmp/stablediffusion/sunflower-2.pn...
I totally ignore that recommendation and sometimes have a happily working system and sometimes it's a wreck. In particular, the upstream Linux driver (what you have anyway) and the dkms built one are sometimes rather divergent. Llvm in rocm has different behaviour to the upstream llvm too.
I'm cautiously optimistic that Debian is packaging will end up as turn key just-works. Part of that is persuading AND developers that breaking ABI has consequences, which it kind of doesn't on the install all of rocm together model.
Right now, I use Debian with the same kernel version that the ROCm recommended Ubuntu ships with and build the dkms driver source intended for Ubuntu on it. That be the wrong thing to do but also totally solid in ymmv fashion.
I'm thinking to begin developing on Linux but love to know what is the current strategy to develop on or for Linux.
What I'm trying to say is of WSL2 and all the work Microsoft is doing with Linux operating system has any other bigger purpose beyond docker or container
I used to use Linux as a primary OS and sometimes dual boot in Windows. However, I have since transitioned to using Windows directly and WSL2. It is seamless overall.
There are a few edge cases, e.g. it was as little annoying to set up LUKS encryption and it's a lot easier to just use BitLocker. Other minor edge cases with permission of volumes that are mounted from Windows into Linux.
Further, there are some nuances around networking, especially if you are running docker in WSL2.
Once you get over these hiccups, it's quite comfortable.
I played with WSL2 and it's interesting to see Windows can directly access Wsl2 Ubuntu directory.
I'm sure it has limitation such as once I tried LLM model and WSL2 lacked support for AMDGPU, therefore it did not have any GPU acceleration.
I am doing UI,backend API services, ML research, ML services
Honestly, the Linux ecosystem is more than mature - if not better - for most of the things you need as a developer imo. The edge cases are usually when you'd need a: 1) A specific proprietary toolchain/IDE for some microcontroller. 2) Windows specific C++/.Net Projects that didn't work well on VS Code.
The shell and desktop environment (for me, KDE Plasma + Yakuake drop down terminal + Dolphin File Manager + KDE Connect) offers a way more superior development experience than anything I found on Windows.
I have been told that Ubuntu based distros are a bit finnicky these days. So, here's the obligatory: "I use Arch btw."
Of course you can hose your system on Arch pretty easily, but it's not like it's difficult to hose your system with Debian or Ubuntu variants. The sneakier thing is that on Debian/Ubuntu variants, there will often be tools, be they third party repositories or even utilities included with the OS, that will make certain bespoke changes to your system for the sake of e.g. installing proprietary drivers, and the combination of those changes will leave you with a snowflake system that may or may not explode when you go to do something complicated (like upgrade major versions.) Whereas with Arch, _a lot_ of the things that you can do that will hose the system will _feel_ like things that will hose the system, and a lot of the scary-looking problems are surprisingly easy to resolve.
Arch shouldn't be any better since obviously, you still need e.g. proprietary drivers or non-standard packages, right? But in my experience, using the AUR to grab those things yields a very good success rate with a low chance of impacting system stability. It's not _zero_ - you can definitely run into problems - but the AUR is generally more reliable for me than third-party dpkg repos have been, and this seems to still be true today.
I wrote a huge wall of text below that nobody will read because it's too long and meandering, but I really think it's time to push for immutable Linux. Arch is oddly stable for what it is, and in my opinion it's a testament to the power of simplicity. But, SteamOS and Fedora Silverblue are starting to show that using modern Linux technologies, you can bring the benefits of immutable operating system design to the Linux desktop today. There's obviously much work to be done, but in my mind the tipping point is either here or coming very soon.
Amen. You hit the nail when it comes to AUR vs. ppa. Steam OS was also going to bring in support for Nix packages, so definitely want to see how immutable desktop + nixpackages end up changing Linux Desktop.
And I'd like to share an anecdote - I installed Pop!_OS in a Samsung Galaxy Book3 Ultra (a combination of weird names) 2 days ago and I'm pleasantly surprised with everything that worked out of the box. The NVIDIA driver works, switching to Intel or NVIDIA works out of the box with zero issues. They even have a nifty command line tool, system76-power.
There are some rough edges. For example, the notebook's speakers do not work, but audio works via bluetooth or connected to a TV. Apparently, Linux Kernel 6.8 should solve that, and I could install it via PPA, but... I'm too afraid to do it. I got old, my primary job is no longer as a dev, and I don't want to spend all that time fixing a Linux system. The Arch Wiki for NVIDIA is incredibly detailed, but Pop!_OS... solves that out of the box.
I feel you. When i got a new laptop, even i couldn't be bothered with Arch install process, so went with Manjaro - just so i wouldn't have to worry about hardware compatibility and still get latest enough kernels.
For a new switching user choosing stable Ubuntu-based distro is the best choice. It has the best software support, the best informational support (distro articles, googling problems), the best quality (because of the popularity).
I am using Mint since 2014 and I am still using the same install of 9 years ago. I update software packages whenever. With big OS updates that come twice a year I wait just in case for 2 extra months and update to that new version. 9 years and counting.
And now imagine an Arch user and how many times something would break during 9 years of operation... It is very telling when discussing this topic and Arch users use arguments like "Arch doesn't really break that much" without truly realizing what others hear and notice. "THAT much", Karl!..
lol.
I feel bad responding with just "lol", but I hope you do actually see how this is quite silly. There is indeed no simple mathematical or statistical fact that says running out of date software with downstream custom patches is more reliable than running the latest version from the developer. If anything, it can cause a lot of difficult-to-detect stability issues that may not impact all users or all use cases.
Not only that, but a big problem with "stable" distros is that most people don't want to use e.g. OBS from 2 years ago, so they need some way to run the latest software. Flatpak or Snap? Sure, but then how do you use e.g. OBS plugins? Suddenly, you are back at "OK, maybe I need a PPA" at which point you need to do things that will inevitably make your system less stable because you are now running somewhat of a "snowflake" configuration that gets more unique every time you add a new PPA or non-trivial modification. Whereas in Arch, you just install it from the package manager, or at worst, AUR, which due to the vastly simpler package management system, is a lot less prone to breakage.
Let me summarize:
- I don't agree that rolling release distros are inherently less stable overall. Stable distros are "tested" but blood, sweat and tears can only go so far. For some really compelling evidence, please ask the Linux kernel folks how their 'LTS' project went. As it turns out, maintaining LTS software is non-trivial in and of itself, and it introduces new problems that did not exist originally.
- Even if rolling release distros were inherently less stable, the reduced need to rely on third-party repositories, packages and even out-of-package-manager installations due to the more up to date packages would offset a lot of that, considering a large part of the problem with Debian/Ubuntu is also just simply that installations get borked too easily.
- Even if that weren't true, Arch's much simpler package management has less of a tendency to get tangled up in impossible-to-resolve dependency issues. Part of this is due to the nature of the AUR vs third-party PPAs, and part of it is just that it's literally much simpler overall, so there's less to go wrong. (Arch packages often err on the side of being less modular, which has its downsides but it certainly simplifies many things.)
- Even if that weren't true, my general experience running mainline Linux and bleeding edge packages for years as my primary operating environment suggests it's actually not very common to be hit with rolling release breakages. Usually people who release software do not just blindly release broken shit. Yes sure, regressions happen, this is a fact of life, but that doesn't mean that you're better off using old versions of things, there's definitely a balance, you don't get to have the benefits of having the newest fixes and never having a regression. In my opinion, frontloading the pain of rather occasional regressions is well worth getting the benefits of the new versions sooner, especially if rolling back is easy. And that, among other reasons, is Exactly why immutable Linux is the future.
P.S.: Yes, "that much" is a perfectly valid thing to say. I have spent many, many, many hours debugging Debian and Ubuntu issues, so the problem isn't that those distros never break and Arch does. The problem is that all distros break and Debian and Ubuntu installs mysteriously, despite being stable distros, absolutely seem to have the most trouble, especially during upgrades. And zero isn't an option. Windows installations break too. Sometimes, especially recently, every Windows installation breaks at once.
I don't think they want 0 breaks but one thing those stable distros can do is isolate you from breaking changes. Think semantic versioning but for package lists. Especially if there are packages that don't play well with each other. iirc gtk2 vs. gtk3 was one such thing? There's a reason a lot of people still use debian etc.. on servers. It is that they just chose a configuration that works and they don't care for/about any breaking updates - as long as they get stable security patches etc..
There is a reason why Release Candidates exist. There is a reason why testing exists and why, despite that, there will always be things not working for the first time, that would require bug-fixing releases.
And yes, Stable releases being more stable than bleeding-age, is a mathematical/statistical fact. You can prove it yourself with the magic of a spreadsheet: - Make a graph that grows X1% on average. This indicates the quality of the software (it is getting better, after all; otherwise there would be no point in updating). But that is the average growth. In practice it is random; daily it can grow or decline by X2, pretty severely as well, but on average it grows. - You can play with X1 and X2 numbers. - Calculate percentages that the update will be better than the previous state for both cases: (1) update 365 times, (2) update 2 times - also take into account that dropping the quality is not equal in strength to growing by the same number. There is a good reason why people are risk-aversed. Quality being dropped means all kinds of trouble and should be calculated as negatively impacting the following X3 (play with this number) days.
If you do that, you will mathematically prove me right. :-)
I am using same Arch install on my desktop for the last 7 or so years. So far, the only time it broke for me was during the whole pulseaudio -> pipewire jump. Half of that is probably due to my pulseaudio tweaks. There's always Manjaro (I use this on my laptop) or other semi rolling release distros too. You are still subscribed to the same repositories but your updates lag by a month or two and would be released to you only after they are happy with their tests.
The main reason I wasn't recommending Ubuntu is because of it's recent annoyances. I recently recommended Ubuntu to a Linux newbie, but after seeing him deal with issues of having to add ppas for some of the toolchains he needed, deal with snap headaches for steam, How gnome takes away certain basic features like being able to edit the path in it's open dialogs, it seemed like there are better options.
Now, I agree, Ubuntu seems to have got a bit to sh*t.
I guess nowadays I would recommend Cinnamon to everybody. Or perhaps PopOs for people from Macs, not sure?
Everything about Windows for software development is worse in my experience. From basics like text rendering to more complex things like managing toolchains and project dependencies. I've done a bit of work porting software from MacOS and Linux to Windows in order to support Windows users (because there are still many) but holy shit, it's a nightmare.
I don't know yet if WSL is a death rattle for software development on Windows, but it feels like one.
I use it everyday, it's Windows only, and I didn't find an alternative as easy and "free" (it requires a Nvidia card, but there is no subscription. Unlike, say, Krisp.ai)
I think it's more of an embrace/extend/extinguish thing.
I'm always good to braindump my thoughts so here goes.
Hardware:
If you want to run Linux as a primary desktop OS, please think about your hardware choices. There's a surprising amount of tradeoffs, and while the Linux desktop is generally a mess, I'd say this roughly defines people who wind up loving it versus hating it. Even in ways you could never realize, it will just impact your entire experience.
- GPUs: If you go with AMD, you have to be aware that choosing the latest cards may not be the best choice. The software stack is rarely ever fully ready for a launch, and if you are not tracking the latest stable kernel it will take even longer to get kernel driver updates. That said, I strongly recommend AMD in general as AMD cards have been very reliable for me and have good drivers across the board once they're ready. NVIDIA has the advantage that you can pretty much get a card at launch and run it on Linux, one of the notable advantages of them staying out of mainline; on the other hand, if you want to keep up with the latest Linux kernels, you may find yourself unable to get the NVIDIA driver working sometimes, as it does not always work on mainline. Further, NVIDIA's drivers still have some issues on Wayland, although it's not nearly as bad as it used to be, though for compatibility reasons I would still not strongly recommend it.
- NICs: Intel NICs generally work pretty good. They're not always perfect but they almost always have good mainline Linux support and reliability issues are definitely an exception case. WiFi is always a bit more persnickety, but things have gotten a lot better. The particular on-board WiFi used by my current motherboard is actually not generally recommended (A MEDIATEK MT7922), but I've been tracking mainline Linux and the only issues I've run into were unrelated. (There was an annoying issue with 6.6.5, it was fixed by 6.6.6 a few days later.)
- Audio: Believe it or not, Linux audio has gotten very compelling lately, after decades of being probably the funniest joke about Linux. Be skeptical if you must, but Pipewire + Wireplumber is probably about as good as Linux audio has been, and has compelling pros and cons when compared even with commercial OSes. Bluetooth will probably work out of the box, and getting low latency "Pro" audio routing features without breaking your desktop audio no longer requires a bunch of specialized knowledge; nominally, it's a dropdown in your audio settings. That having been said... You may face some issues, especially with cutting-edge motherboards and laptops. In general, Linux does not adequately support laptops that need a lot of software DSP in the driver to sound good. In addition, some of the more advanced audio codec setups may not work correctly out of the box. This is NOT the case for the majority of desktop motherboards, but it is worth googling to see if you can find someone reporting that it works.
Software:
There is a lot of heated debates about what the best distro and desktop environments are. Normally, I'd recommend sticking with a stalwart like Debian, but in this changing environment I'm really not sure what I'd personally recommend. I run NixOS, which I think is a fantastic operating system for people who know they need it, but like Linux itself, I would not recommend it to anyone unless they're relatively sure it's something that solves a problem they actually have. I pretty much knew immediately I wanted it when I heard about it.
But even when not considering NixOS, I strongly would recommend trying to find an "immutable" Linux distro that will work for you. This is because immutable Linux has proven to be a much better model for desktop usage that is far more robust and easier to reason about. A good example would be Fedora Silverblue. There was a nice looking project on here recently that may be of particular interest to developers, based on Silverblue, called Bluefin:
I can't necessarily personally recommend it, but the concept seems solid. Recommending a random recent project is a little risky, but it at least has the benefit that it's essentially just a spin of Silverblue that can be reverted back to plain old Silverblue safely as far as I know.
My personal choice for a desktop environment is none. I use SwayWM with a bunch of bespoke configuration. It's... ugly[1]. But it works, and for me, it works quite well.
But that is a horrible recommendation for someone who is trying to get into Linux by any measure. Instead, I recommend KDE. KDE has a fair amount of stability issues, but what I can say about KDE is that it's got a great balance of customization, features and polish for a desktop system, and I think that the current iteration of KDE is going to age well into the future. We'll see how that thought holds up, but I have my hopes.
If you are using an AMD card, I can also say that running KDE/Plasma Desktop in Wayland mode is a pretty good experience, if you can. It does come with some challenges and downsides, but there are some positives, especially if you like high DPI or high refresh rate screens, or you want to use OBS to capture or stream the screen, as it will be a lot more efficient using Pipewire on Wayland. At the very least, it's not a huge commitment since you can log out and jump back in with an X11 session through your display manager if anything goes too wrong.
P.S.: While I don't immediately recommend anyone to go and try NixOS, I can't say the same about Nix. Go and install Nix on your macOS or Linux machines! It's a great tool to have and is generally not obtrusive on a machine. Especially see the `nix develop` command. It drops you in a shell with the development tools used to build a package, so I can do say, `nix develop nixpkgs#_86Box` to get all of the dependencies to build the 86Box emulator, and just start building it with CMake.
[1]: https://github.com/jchv/nixos-config/tree/master/modules/pro... - In case anyone is curious about a "real world" Sway setup, this certainly is one. In the back of my mind, I'm secretly hoping someone will suggest improvements. :)
Unless you're working on something that is heavily Windows-specific (legacy .NET Framework, Win32, Visual C++, etc.) that requires Visual Studio and Windows as a runtime developing on Linux is a much nicer experience.
Can anyone add some context beyond the post as to what this is?
Is it memory mapped data buffer and circular command buffers?
This driver both provides necessary interface in place of an ARM core talking direct over AXI buses, and the glue logic to enable use of XRT.