Debian, Arch and Gentoo have ROCm built for consumer GPUs. Thus so do their derivatives. Anything gfx9 or later is likely to be fine and gfx8 has a decent chance of working. The https://github.com/ROCm/ROCm source has build scripts these days.
At least some of the internal developers largely work on consumer hardware. It's not as solid as the enterprise gear but it's also very cheap so overall that seems reasonable to me. I'm using a pair of 6900XT, with a pair of VII's in a backup machine.
For turn key proprietary stuff where you really like the happy path foreseen by your vendor, in classic mainframe style, team green is who you want.
> For turn key proprietary stuff where you really like the happy path foreseen by your vendor
there really was no way for AMD to foresee that people might want to run GPGPU workloads on their polaris cards? isn't that a little counterfactual to the whole OpenCL and HSA Framework push predating that?
Example: it's not that things like Bolt didn't exist to try and compete with Thrust... it's that the NVIDIA one has had three updates in the last month and Bolt was last updated 10 years ago.
You're literally reframing "having working runtime and framework support for your hardware" as being some proprietary turnkey luxury for users, as well as an unforeseeable eventuality for AMD. It wasn't a development priority, but users do like to actually build code that works etc.
That's why you got kicked to the curb by Blender - your OpenCL wasn't stable even after years of work from them and you. That's why you got kicked to the curb by Octane - your Vulkan Compute support wasn't stable enough to even compile their code successfully. That's the story that's related by richg42 about your OpenGL driver implementation too - that it's just paper features and resume-driven development by developers 10 years departed all the way down.
The issues discussed by geohotz aren't new, and they aren't limited to ROCm or deep learning in general. This is, broadly speaking, the same level of quality that AMD has applied to all its software for decades. And the social-media "red team" loyalism strategy doesn't really work here, you can't push this into "AMD drivers have been good for like 10 years now!!!" fervor when the understanding of the problems are that broad and that collectively shared and understood. Every GPGPU developer who's tried has bounced off this AMD experience for literally an entire generation running now. The shared collective experience is that AMD is not serious in the field, and it's difficult to believe it's a good-faith change and interest in advancing the field rather than just a cashgrab.
It's also completely foreseeable that users want broad, official support for all their architectures, and not one or two specifics etc. Like these aren't mysteries that AMD just accidentally forgot about, etc. They're basic asks that you are framing as "turnkey proprietary stuff", like a working opencl runtime or a working spir-v compiler.
What was it linus said about the experience of working with NVIDIA? That's been the experience of the GPGPU community working with AMD, for decades. Shit is broken and doesn't work, and there's no interest in making it otherwise. And the only thing that changed it is a cashgrab, and a working compiler/runtime is still "turnkey proprietary stuff" they have to be arm-twisted into doing by Literally Being Put On Blast By Geohotz Until It's Fixed. "Fuck you, AMD" is a sentiment that there is very valid reasons to feel given the amount of needless suffering you have generated - but we just don't do that to red team, do we?
But you guys have been more intransigent about just supporting GPGPU, no matter what framework, please just pick one, get serious and start working already than NVIDIA ever was about wayland etc. You've blown decades just refusing to ever shit or get off the pot (without even giving enough documentation for the community to just do it themselves). And that's not an exaggeration - I bounced off the AMD stack in 2012, and it wasn't a new problem then either. It's too late for "we didn't know people wanted a working runtime or to develop on gaming cards" to work as an excuse, after decades of overt willing neglect it's just patronizing.
Again, sorry, this is ranty, it's not that I'm upset at you personally etc, but like, my advice as a corporate posture here is don't go looking for a ticker-tape parade for finally delivering a working runtime that you've literally been advertising support for for more than a decade like it's some favor to the community. These aren't "proprietary turnkey features" they're literally the basics of the specs you're advertising compliance with, and it's not even just one it's like 4+ different APIs that have this problem with you guys that has been widely known, discussed in tech blogs etc for more than a decade (richg42). I've been saying it for a long time, so has everyone else who's ever interacted with AMD hardware in the GPGPU space. Nobody there cared until it was a cashgrab, actually half the time you get the AMD fan there to tell you the drivers have been good for a decade now (AMD cannot fail, only be failed). It's frustrating. You've poisoned the well with generations of developers, with decades of corporate obstinance that would make NVIDIA blush, please at least have a little contrition about the whole experience and the feelings on the other side here.
You're saying interesting things here. It's not my perspective but I can see how you'd arrive at it. Worth noting that I'm an engineer writing from personal experience, the corporate posture might be quite divergent from this.
I think Cuda's GPU offloading model is very boring. An x64 thread occasionally pushes a large blob of work into a stream and sometime later finds out if it worked. That does however work robustly, provided you don't do anything strange from within the kernel. In particular allocating memory on the host from within the kernel deadlocks the kernel unless you do awkward things with shuffling streams. More ambitious things like spawning a kernel from a kernel just aren't available - there's only a hobbled nested lifetime thing available. The volta threading model is not boring but it is terrible, see https://stackoverflow.com/questions/64775620/cuda-sync-funct...
HSA puts the x64 cores and the gpu cores on close to equal footing. Spawning a kernel from a kernel is totally fine and looks very like spawning one from the host. Everything is correctly thread safe so calling mmap from within a kernel doesn't deadlock things. You can program the machine as a large cluster of independent cores passing messages to one another. For the raw plumbing, I wrote https://github.com/jonchesterfield/hostrpc. That can do things like have an nvidia card call a function on an amd one. That's the GPU programming model I care about - not passing blobs of floating point math onto some accelerator card, I want distributed graph algorithms where the same C++ runs on different architectures transparently to the application. HSA lends itself to that better than Cuda does. But it is rather bring your own code.
That is, I think the more general architecture amdgpu is shipping is better than the specialised one cuda implements, despite the developer experience being rather gnarlier. I can't express the things I want to on nvptx at all so it doesn't matter much that simpler things would work more reliably.
Maybe more relevant to your experience, I can offer some insight into the state of play at AMD recently and some educated guesses at the earlier state. ATI didn't do compute as far as I know. Cuda was announced in 2006, same year AMD acquired ATI. Intel Core 2 was also 2006 and I remember that one as the event that stopped everyone buying AMD processors. Must have been an interesting year to be in semiconductors, was before my time. So in the year cuda appears, ATI is struggling enough to be acquired, AMD mortgages itself to the limit to make the acquisition and Intel obsoletes AMD's main product.
I would guess that ~2007 marked the beginning of the really bad times for AMD. Even if they could guess what cuda would become they were in no position to do anything about it. There is scar tissue still evident from that experience. In particular, the games console being the breadwinner for years can be seen in some of the hardware decisions, and I've had an argument with someone whose stance was that semi-custom doesn't need a feature so we shouldn't do it.
What turned the corner is the DoE labs being badly burned by reliance on a single vendor for HPC. AMD proposed a machine which looks suspiciously like a lot of games consoles with the power budget turned way up and won the Frontier bid with it. That then came with a bunch of money to try to write some software to run on it which in a literal sense created the job opening I filled five years back. Intel also proposed a machine which they've done a hilariously poor job of shipping. So now AMD has built a software stack which was razor focused on getting the DoE labs to sign the cheques for functionally adequate on the HPC machines. That's probably the root of things like the approved hardware list for ROCm containing the cards sold to supercomputers and not so much the other ones.
It turns out there's a huge market opportunity for generative AI. That's not totally what the architecture was meant to do but whatever, it likes memory bandwidth and the amdgpu arch does do memory bandwidth properly. The rough play for that seems to be to hire a bunch of engineers and buy a bunch of compiler consultancies and hope working software emerges from that process, which in fairness does seem to be happening. The ROCm stack is irritating today but it's a whole different level of QoI relative to before the Frontier bring up.
Note that there's no apology nor contrition here. AMD was in a fight to survive for ages and rightly believed that R&D on GPU compute was a luxury expense. When a budget to make it work on HPC appeared it was spent on said HPC for reasonable fear that they wouldn't make the stage gates otherwise. I think they've done the right thing from a top level commercial perspective for a long time - the ATI and Xilinx merges in particular look great.
Most of my colleagues think the ROCm stack works well. They use the approved Ubuntu kernel and a set of ROCm libraries that passed release testing to iterate on their part of the stack. I suspect most people who treat the kernel version and driver installation directions as important have a good experience. I'm closer to the HN stereotype in that I stubbornly ignore the binary ROCm release and work with llvm upstream and the linux driver in whatever state they happen to be in, using gaming cards which usually aren't on the supported list. I don't usually have a good time but it has definitely got better over the years.
I'm happy with my bet on AMD over Nvidia, despite the current stock price behaviour making it a serious financial misstep. I believe Lisa knows what she's doing and that the software stack is moving in the right direction at a sufficient pace.
Although "official" / validated support^ is only for PRO W6800/V620 for RDNA2 and RDNA3 RX 7900's for consumer. Based on lots of reports you can probably just HSA_OVERRIDE_GFX_VERSION override for other RDNA2/3 cards and it'll probably just work. I can get GPU-accelerate ROCm for LLM inferencing on my Radeon 780M iGPU for example w/ ROCm 6.0 and HSA_OVERRIDE_GFX_VERSION=11.0.0
(In the past some people also built custom versions of ROCm for older architectures (eg ROC_ENABLE_PRE_VEGA=1) but I have no idea if those work still or not.)
Debian, Arch and Gentoo have ROCm built for consumer GPUs. Thus so do their derivatives. Anything gfx9 or later is likely to be fine and gfx8 has a decent chance of working. The https://github.com/ROCm/ROCm source has build scripts these days.
At least some of the internal developers largely work on consumer hardware. It's not as solid as the enterprise gear but it's also very cheap so overall that seems reasonable to me. I'm using a pair of 6900XT, with a pair of VII's in a backup machine.
For turn key proprietary stuff where you really like the happy path foreseen by your vendor, in classic mainframe style, team green is who you want.
> For turn key proprietary stuff where you really like the happy path foreseen by your vendor
there really was no way for AMD to foresee that people might want to run GPGPU workloads on their polaris cards? isn't that a little counterfactual to the whole OpenCL and HSA Framework push predating that?
Example: it's not that things like Bolt didn't exist to try and compete with Thrust... it's that the NVIDIA one has had three updates in the last month and Bolt was last updated 10 years ago.
You're literally reframing "having working runtime and framework support for your hardware" as being some proprietary turnkey luxury for users, as well as an unforeseeable eventuality for AMD. It wasn't a development priority, but users do like to actually build code that works etc.
That's why you got kicked to the curb by Blender - your OpenCL wasn't stable even after years of work from them and you. That's why you got kicked to the curb by Octane - your Vulkan Compute support wasn't stable enough to even compile their code successfully. That's the story that's related by richg42 about your OpenGL driver implementation too - that it's just paper features and resume-driven development by developers 10 years departed all the way down.
The issues discussed by geohotz aren't new, and they aren't limited to ROCm or deep learning in general. This is, broadly speaking, the same level of quality that AMD has applied to all its software for decades. And the social-media "red team" loyalism strategy doesn't really work here, you can't push this into "AMD drivers have been good for like 10 years now!!!" fervor when the understanding of the problems are that broad and that collectively shared and understood. Every GPGPU developer who's tried has bounced off this AMD experience for literally an entire generation running now. The shared collective experience is that AMD is not serious in the field, and it's difficult to believe it's a good-faith change and interest in advancing the field rather than just a cashgrab.
It's also completely foreseeable that users want broad, official support for all their architectures, and not one or two specifics etc. Like these aren't mysteries that AMD just accidentally forgot about, etc. They're basic asks that you are framing as "turnkey proprietary stuff", like a working opencl runtime or a working spir-v compiler.
What was it linus said about the experience of working with NVIDIA? That's been the experience of the GPGPU community working with AMD, for decades. Shit is broken and doesn't work, and there's no interest in making it otherwise. And the only thing that changed it is a cashgrab, and a working compiler/runtime is still "turnkey proprietary stuff" they have to be arm-twisted into doing by Literally Being Put On Blast By Geohotz Until It's Fixed. "Fuck you, AMD" is a sentiment that there is very valid reasons to feel given the amount of needless suffering you have generated - but we just don't do that to red team, do we?
But you guys have been more intransigent about just supporting GPGPU, no matter what framework, please just pick one, get serious and start working already than NVIDIA ever was about wayland etc. You've blown decades just refusing to ever shit or get off the pot (without even giving enough documentation for the community to just do it themselves). And that's not an exaggeration - I bounced off the AMD stack in 2012, and it wasn't a new problem then either. It's too late for "we didn't know people wanted a working runtime or to develop on gaming cards" to work as an excuse, after decades of overt willing neglect it's just patronizing.
Again, sorry, this is ranty, it's not that I'm upset at you personally etc, but like, my advice as a corporate posture here is don't go looking for a ticker-tape parade for finally delivering a working runtime that you've literally been advertising support for for more than a decade like it's some favor to the community. These aren't "proprietary turnkey features" they're literally the basics of the specs you're advertising compliance with, and it's not even just one it's like 4+ different APIs that have this problem with you guys that has been widely known, discussed in tech blogs etc for more than a decade (richg42). I've been saying it for a long time, so has everyone else who's ever interacted with AMD hardware in the GPGPU space. Nobody there cared until it was a cashgrab, actually half the time you get the AMD fan there to tell you the drivers have been good for a decade now (AMD cannot fail, only be failed). It's frustrating. You've poisoned the well with generations of developers, with decades of corporate obstinance that would make NVIDIA blush, please at least have a little contrition about the whole experience and the feelings on the other side here.
You're saying interesting things here. It's not my perspective but I can see how you'd arrive at it. Worth noting that I'm an engineer writing from personal experience, the corporate posture might be quite divergent from this.
I think Cuda's GPU offloading model is very boring. An x64 thread occasionally pushes a large blob of work into a stream and sometime later finds out if it worked. That does however work robustly, provided you don't do anything strange from within the kernel. In particular allocating memory on the host from within the kernel deadlocks the kernel unless you do awkward things with shuffling streams. More ambitious things like spawning a kernel from a kernel just aren't available - there's only a hobbled nested lifetime thing available. The volta threading model is not boring but it is terrible, see https://stackoverflow.com/questions/64775620/cuda-sync-funct...
HSA puts the x64 cores and the gpu cores on close to equal footing. Spawning a kernel from a kernel is totally fine and looks very like spawning one from the host. Everything is correctly thread safe so calling mmap from within a kernel doesn't deadlock things. You can program the machine as a large cluster of independent cores passing messages to one another. For the raw plumbing, I wrote https://github.com/jonchesterfield/hostrpc. That can do things like have an nvidia card call a function on an amd one. That's the GPU programming model I care about - not passing blobs of floating point math onto some accelerator card, I want distributed graph algorithms where the same C++ runs on different architectures transparently to the application. HSA lends itself to that better than Cuda does. But it is rather bring your own code.
That is, I think the more general architecture amdgpu is shipping is better than the specialised one cuda implements, despite the developer experience being rather gnarlier. I can't express the things I want to on nvptx at all so it doesn't matter much that simpler things would work more reliably.
Maybe more relevant to your experience, I can offer some insight into the state of play at AMD recently and some educated guesses at the earlier state. ATI didn't do compute as far as I know. Cuda was announced in 2006, same year AMD acquired ATI. Intel Core 2 was also 2006 and I remember that one as the event that stopped everyone buying AMD processors. Must have been an interesting year to be in semiconductors, was before my time. So in the year cuda appears, ATI is struggling enough to be acquired, AMD mortgages itself to the limit to make the acquisition and Intel obsoletes AMD's main product.
I would guess that ~2007 marked the beginning of the really bad times for AMD. Even if they could guess what cuda would become they were in no position to do anything about it. There is scar tissue still evident from that experience. In particular, the games console being the breadwinner for years can be seen in some of the hardware decisions, and I've had an argument with someone whose stance was that semi-custom doesn't need a feature so we shouldn't do it.
What turned the corner is the DoE labs being badly burned by reliance on a single vendor for HPC. AMD proposed a machine which looks suspiciously like a lot of games consoles with the power budget turned way up and won the Frontier bid with it. That then came with a bunch of money to try to write some software to run on it which in a literal sense created the job opening I filled five years back. Intel also proposed a machine which they've done a hilariously poor job of shipping. So now AMD has built a software stack which was razor focused on getting the DoE labs to sign the cheques for functionally adequate on the HPC machines. That's probably the root of things like the approved hardware list for ROCm containing the cards sold to supercomputers and not so much the other ones.
It turns out there's a huge market opportunity for generative AI. That's not totally what the architecture was meant to do but whatever, it likes memory bandwidth and the amdgpu arch does do memory bandwidth properly. The rough play for that seems to be to hire a bunch of engineers and buy a bunch of compiler consultancies and hope working software emerges from that process, which in fairness does seem to be happening. The ROCm stack is irritating today but it's a whole different level of QoI relative to before the Frontier bring up.
Note that there's no apology nor contrition here. AMD was in a fight to survive for ages and rightly believed that R&D on GPU compute was a luxury expense. When a budget to make it work on HPC appeared it was spent on said HPC for reasonable fear that they wouldn't make the stage gates otherwise. I think they've done the right thing from a top level commercial perspective for a long time - the ATI and Xilinx merges in particular look great.
Most of my colleagues think the ROCm stack works well. They use the approved Ubuntu kernel and a set of ROCm libraries that passed release testing to iterate on their part of the stack. I suspect most people who treat the kernel version and driver installation directions as important have a good experience. I'm closer to the HN stereotype in that I stubbornly ignore the binary ROCm release and work with llvm upstream and the linux driver in whatever state they happen to be in, using gaming cards which usually aren't on the supported list. I don't usually have a good time but it has definitely got better over the years.
I'm happy with my bet on AMD over Nvidia, despite the current stock price behaviour making it a serious financial misstep. I believe Lisa knows what she's doing and that the software stack is moving in the right direction at a sufficient pace.
7 replies →
ROCm 6.0 and 6.1 list RDNA3 (gfx1100) and RDNA2 (gfx1030) in their supported architectures list: https://rocm.docs.amd.com/en/latest/compatibility/compatibil...
Although "official" / validated support^ is only for PRO W6800/V620 for RDNA2 and RDNA3 RX 7900's for consumer. Based on lots of reports you can probably just HSA_OVERRIDE_GFX_VERSION override for other RDNA2/3 cards and it'll probably just work. I can get GPU-accelerate ROCm for LLM inferencing on my Radeon 780M iGPU for example w/ ROCm 6.0 and HSA_OVERRIDE_GFX_VERSION=11.0.0
(In the past some people also built custom versions of ROCm for older architectures (eg ROC_ENABLE_PRE_VEGA=1) but I have no idea if those work still or not.)
^ https://rocm.docs.amd.com/projects/install-on-linux/en/lates...