Lczero Neural Network Chess Benchmarks With OpenCL Radeon vs. NVIDIA

[phoronix]

Phoronix: Lczero Neural Network Chess Benchmarks With OpenCL Radeon vs. NVIDIA

Yesterday I posted a number of Lczero chess engine benchmarks on NVIDIA GPUs using its OpenCL back-end as well as its CUDA+cuDNN back-end, which offered massive performance gains compared to CL on the many tested NVIDIA GPUs. With the CUDA+cuDNN code performing so much better than OpenCL, some wondered whether NVIDIA was intentionally gimping their OpenCL performance. Well, here are results side-by-side now with Radeon GPUs on OpenCL...

http://www.phoronix.com/scan.php?page=news_item&px=Lczero-OpenCL-Radeon-NVIDIA

[gorgone]

the RX Vega 64 is slower than RX 590
this code must be a masterpiece^^

[ObiWan]

And rocm2 beeing slow, this is with a vega56 and the amdgpu-pro opencl backend:
https://openbenchmarking.org/result/...SK-VEGA5685245

[leidola]

That's a factor two. Rocm opencl really needs some tuning.

[ankan]

For CPU backend, I think you can link to intel mkl-dnn lib (instead of OpenBLAS) to get much better performance.
Search for mkl-dnn in github
Also, I assume you are already running it with enough no. of threads (e.g: --threads=32).

[Marc Driftmeyer]

Whoever wrote their OpenCL stack should be shot. It reminds me of how piss poor Blender's monolithic stack was before AMD rewrote it into a split stack and threaded it properly.

[Marc Driftmeyer]

And rocm2 beeing slow, this is with a vega56 and the amdgpu-pro opencl backend:
https://openbenchmarking.org/result/...SK-VEGA5685245

It isn't just ROCm. Most of these slowness comes from the client's poor coding and knowledge of OpenCL as well.

[zxy_thf]

It isn't just ROCm. Most of these slowness comes from the client's poor coding and knowledge of OpenCL as well.

Additionally I don't even think it is a good idea to use OpenCL directly rather than building the AI engine on top of pytorch/tensorflow.

[alcalde]

Additionally I don't even think it is a good idea to use OpenCL directly rather than building the AI engine on top of pytorch/tensorflow.

I thought neither TensorFlow nor PyTorch worked with OpenCL?

[kohlerm]

Thanks, very interesting test!
So getting a 2060 is now probably the way to go to get a really strong chess engine. lczero just got the second place at the TCEC computer chess tournament, behind stockfish.
But with a 2060 you would need to get quite some expensive hardware to make Stockfish a competitor. Would need to find some numbers for stockfish speed depending on the number of cores, but currently at a ratio of 1:1000 in nodes per second lczero is most likely stronger than stockfish. That being said stockfish would be still very useful for deep tactical analysis problems.