rdna 2 support?

[bernharl]

Will ROCM support the rdna 2 architecture when it is released in the future? I and a lot of others would love to be able to do some machine learning research on my personal computer and I would love to switch over to AMD because of their open sourced Linux drivers, but I'm not sure if I could actually make that switch if I can't replace the CUDA support in my Geforce card.

I think the lack of Navi support is already a missed opportunity for AMD, but it sort of makes sense not to spend a massive amount of time supporting a card that isn't really flagship-tier. However, if the rdna 2 series of cards includes cards significantly more powerful than the 5700XT this wouldn't be the case anymore.

[fishhf]

AMD should support both RDNA 1 and 2. Otherwise people such as me will hop back to Nvidia after getting burned by 5700XT.

[bernharl]

I agree, ideally it would be like CUDA where any Geforce GPU is supported. If AMD actually wants to compete in the ML market this is a must. People just getting into Machine Learning won't buy enterprise AMD GPUs if they can just buy a middle-tier RTX card.

[kevin335200]

Sadly never. AMD has a separate product line called CDNA for computing. I'm afraid RDNA architecture is not suitable for ML by design.

[bernharl]

This has to change for AMD to be a viable option for startups and individual scientists.

[limapedro]

@kevin335200 @bernharl Well at least for Windows users now is possible to use any DX 12 capable device to use tensorflow (Nvidia GPUs, AMD GPUs and APUs). I hope that the release Navi 2 with Rocm support, but now at least I know that I can use AMD GPUs, I hope that tensorflow-directml will get more optimization soon.

[bernharl]

@limapedro Might as well use an Nvidia GPU with Cuda if I were to use proprietary software anyway. The reason I want an AMD GPU is because of the open source Linux driver...

Hopefully with ROCM being open source someone could be able to imlement RDNA2 support themselves is AMD refuses to!

[fishhf]

Tensorflow's XLA backend already have targets that uses LLVM, and RDNA seems to be already in LLVM, I wonder if it's actually a lot of work to create a XLA backend that targets RDNA.
Then both Tensorflow and JAX would run on RDNA.

[cb88]

Apparently 5700xt is working in ROCm to some degree now (like darktable and blender but a bit unstable) ... RDNA2 probably isn't that different from RDNA1 compute wise and neither of them are incapable of compute, they just aren't optimized for it, CDNA is removing extra stuff from the GPU so they can scale to more CUs as well as reduce hardware overhead that exists in GCN and RDNA that ins requried to perform well for graphics (but doesn't really hurt compute either except maybe be a bit less power efficient given ideal code, RDNA can cope better with non ideal code though).

Getting compute working on RNDA is very important for workstation users as AMD does sell the pro cards it just hasn't got the same priority as upcoming cards I think we'll see it finally straighten out once RDNA2 and CDNA are out.

[tuxzz]

If RDNA2 support landed within four weeks after 6800XT sale date and had a similar performance to RTX 3080 on FP32+Tensorflow or other computation task, I would buy a bunch of it for my school laboratory.

[darkbasic]

I wouldn't hold my breath: chances of getting anything in a reasonable timeframe are low but if you're feeling optimistic I would rather bet on clover being first

[cb88]

We basically already have RDNA1 support... it just isn't mainline. So.. I doubt there is a ton else to do for RNDA2.

[fishhf]

Apparently 5700xt is working in ROCm to some degree now (like darktable and blender but a bit unstable)

Darktable and blender should be just using OpenCL. OpenCL support is basically there since day 1 when 5700XT is released.

[MichelBr]

The same here. I'll be looking for a new card/system at the end of the year and the zen 3, together with the rdna2 look great, but I'll only buy it if rocm supports it decently. I recently used rocm on vega with linux and to be honest it works quite well now. The instructions are not too bad and it just worked afterwards.

However, I don't want to fiddle with issues and amd needs to indeed support these cards for non-datacenter-usage (individuals, schools, ...). If they can buy Xilinx, they can also hopefully invest a bit of money in good tensorflow support for invidivuals who work with tensorflow and get more foothold in the ecosystem.

[darkbasic]

ROCm 3.9 just released, still no RDNA support.

[xuhuisheng]

There are codes and branches for gfx103x, Guess RDNA2 maybe supported next year?

https://github.com/ROCmSoftwarePlatform/Tensile/tree/gfx10
ROCm/ROCm-Device-Libs@bf43ad1
ROCm/MIOpen#545