AMD's New RX Radeon VII - NVIDIA RTX 2080 Killer?

[nabildanial]

Does it have A.I. hardware? If not then it's not a 2080 killer as far as chess is concerned.

Afaik, AMD still has no TensorCores on board,
but it remains open how the price per Knps for LC0 will be...

--
Srdja

I think the LC0 code is strictly written for Nvidia cards, so a complete rewrite would be needed for OpenCL.

There are 3 versions of lc0, CUDA, OpenCL and CPU so it's already there. Thing is it's way slower than CUDA.

Srdja is an OpenCL expert, so maybe he can fix it.

Thanks for the Kudos, but i am really a layman in neural networks.

Afaik, GCP wrote Leela Zero initially with an OpenCL backend,
and this was ported to Leela Chess Zero,
then Ankan added an high performant cuDNN backend,
and with the new Nivida RTX series this backend profits from TensorCores.

So i think it is AMDs turn to release some optimized OpenCL NN libraries
and add TensorCores, or a similar technique, to their gpus.

--
Srdja

There's already an effort made for AMD gpus to support CUDA and cuDNN codes, that is already in beta stage. Maybe someone can port it into lc0?

CUDA code to portable C++ converter (HIP):
https://github.com/ROCm-Developer-Tools/HIP

cuDNN to hipDNN:
https://github.com/ROCmSoftwarePlatform/hipDNN

[smatovic]

So i think it is AMDs turn to release some optimized OpenCL NN libraries
and add TensorCores, or a similar technique, to their gpus.

--
Srdja

Here is a NN lib for OpenCL:
https://github.com/hughperkins/DeepCL
I've no idea how good it is.
Or maybe:
https://github.com/plaidml/plaidml
Or maybe:
https://github.com/keras-team/keras

As to your second point:
https://hothardware.com/news/amd-7nm-na ... p-learning

See, you are more expert than me

[smatovic]

There's already an effort made for AMD gpus to support CUDA and cuDNN codes, that is already in beta stage. Maybe someone can port it into lc0?

CUDA code to portable C++ converter (HIP):
https://github.com/ROCm-Developer-Tools/HIP

cuDNN to hipDNN:
https://github.com/ROCmSoftwarePlatform/hipDNN

I don't know about the HIP ecosphere,
but one caveat with ROCm drivers is the lacking support for older AMD GPUs and CPUs without PCIe 3.0

https://rocm.github.io/ROCmInstall.html#supported-gpus
https://rocm.github.io/ROCmInstall.html#supported-cpus

--
Srdja

[nabildanial]

There's already an effort made for AMD gpus to support CUDA and cuDNN codes, that is already in beta stage. Maybe someone can port it into lc0?

CUDA code to portable C++ converter (HIP):
https://github.com/ROCm-Developer-Tools/HIP

cuDNN to hipDNN:
https://github.com/ROCmSoftwarePlatform/hipDNN

I don't know about the HIP ecosphere,
but one caveat with ROCm drivers is the lacking support for older AMD GPUs and CPUs without PCIe 3.0

https://rocm.github.io/ROCmInstall.html#supported-gpus
https://rocm.github.io/ROCmInstall.html#supported-cpus

--
Srdja

Couldn't be any worse than current CUDA support for lc0 where older GPUs without CUDA v9.2 support cannot run it. People will always have OpenCL to fallback on.

EDIT: Didn't realize that ROCm has no Windows support yet, it's a no go then.

[smatovic]

Linux benchmarks are out,
on average it is en par with Nvidia RTX 2080

https://www.phoronix.com/scan.php?page= ... nux&num=11

but i think they messed up somehow the LC0 bench

https://www.phoronix.com/scan.php?page= ... inux&num=9

these nps values look fishy to me.

--
Srdja