Getting Started With PyTorch on AMD GPUs: Community & Partner Talk at PyTorch Conference 2022 

If the new AMD gaming gpus support this, I will be using them for my new computer. Great job!

Really good to see some momentum with ROCm, it needs to catch up to CUDA in e.g. Windows/WSL/Docker support. I really like how the Radeon cards generally have more (often double) VRAM than their price matched Nvidia counterparts. If the software side can catch up that'd be great too.

Haven't tested these new drivers yet because I have Nvidia now, but my last ROCM experience was that it was all a bit janky and hard to use compared to Nvidia. If it manages to work as well as they claim then this is pretty huge. AMD GPU's generally offer a ton more VRAM than Nvidia, so this means that if this works without issues, you can get a lot more performance than Nvidia for cheaper. Not to mention that AMD based mini PC's with APU's will be far easier to work with since they just use X86 instead of Nvidia's Jetson ARM (and Jetson SBC's are extremely expensive for how little performance they offer)

Proud to be part of the team 🙂

The rocm and pytorch-rocm does not run on Windows and you need to wait for months/years to have ROCm to added new cards like 7900XTX and wait linux kernel to add the kernel driver to support, that's why no one use ROCm. AMD need to put it on windows and tract more users to deeplearning using their cards at beginning.

PYTORCH + AMD will be my SALVATION!

Please benchmark Stable Diffusion nvidia vs amd radeon

where can I try AMD Instinct GPU? Searched on Azure cloud, cannot find AMD MI50-250 in their instances. Any suggestions on lab or cloud env? Thank you.

ROCm is not available on Windows, sadly.

can we get this on the rx7900? VRAM is the most important thing for rapidly proof of concepting Ai on a work station. AMD Radeon is positioned to capture this.

Why didn't someone tell about this sooner ? Some dude in chat just casually is like 'Why don't you use ROCM' Granted it's Linux only, I'm usually on Windows. But I'll use Linux if this works :P

Yeh just the change in warp size probably results in a large performance improvement, but I’m curious to know how the differences in GPU cache architecture translate to CUDA programming. There’s no way that CUDA can really use the “infinity cache” or at least there must be differences in how cache benefits/concerns are handled in CUDA code vs GL/Vulkan code. It’s a shame that NVidia has such an iron grip on this stuff because alternative arch/cache designs can improve performance significantly for some workloads. Until OpenMP is ingested through software/system packages it’s not clear that getting off vendor specific code will be possible (probably mencing some acronyms here). However, the simplest approaches in software interfaces that permit code portability b/w compute architectures almost necessarily lose the distinguishing performance benefits offered by any specific architecture. I guess that means that extreme performance gains are still only available to large engineering teams.

sheeeeeeeeeeeeeeeeeeeeeesh

That is all great but GPU compatibility with rocm must be heavily extended. There should be no reason why cards like 6600xt, 6700xt and more importantly 7900xtx aren't supported. Rocm needs to exist on windows as well because right now, you have to resort to directml to build commercial applications for windows, when using AMD cards.

I'll have to buy a new gpu very soon for ML, is it real or not ?

Nothing works, no support for 7800xt, 11.27.2023. Only some OpenCL libs, but it runs at half performance at best.

PLSS AMD needing a computer science degree and kill a dragon just to run language models locally is not fun!!, more support for other gpus would be really appreciated also ROCm for windows when?

I wanna fight this man

Hey AMD, When following will happen? ROCm+pytorch on windows with AMD's Gaming GPUs. ---------------------------- See the Nvidia's stack. Everything is working on windows for them using their gaming GPUs. Yours performance charts are meaningless for regular user because that user is not having instinct gpu in his laptop/PC. At present, It looks like Intel will reach there before AMD.

I am writing this to encourage the authors to make improvements. AMD should fire this guy and his entire team, and start all over. It is unlikely that any freshly formed team could get it to work. I suggest that they start their whole system over, and make something work. Not just for the devs, not just on their charts, and graphs, but normal people. Make a system work, for people on their hardware, and in their own homes. AMD does not have their own working product, and instead they lie. HIP is a bandaid, that really doesn't cover outside of the lab. I'd like to suggest nobody else buy products until they're working. While it be great if self-congratulatory drivel such as this was sufficient help to actually make things work, alas it is not. There's no evidence that any of this stuff works beyond this guy's team. Go and search for it, there's NO working examples... it's a horrid mess. Like for example, yes there's hipify pyton in the torch project - but there's no how to demonstrating it working or how to use it. They haven't bothered to document in their own code, much less show it works. Just because it works for the development team, doesn't mean much else. Their document style leaves a lot to desired like basic function. It appears to be merely pretense of function, it is a false front. It is a shell game, there's really nothing here for common people. And note their platform, while some cases might translate, do not on their own - their own documents say start with working CUDA to add confusion. It isn't its own fully functioning system you can use at home. Their support system woefully lacks in function... it's embarrassing. In this day and age, they should offer a bounty on working demos.

no windows support, so windows users with AMD gpus will still suffer from very bad performance and lack of support in Stable Diffusion, they make me regret i bought AMD GPU instead of Nvidia GPUS for the first time 😡😡😡😡