No announcement yet.

Lczero Neural Network Chess Benchmarks With OpenCL Radeon vs. NVIDIA

  • Filter
  • Time
  • Show
Clear All
new posts

  • #11
    Clover is able to run lczero, too. However, the PTS profile naïvely defaults to requesting maximum CPU threads, which (as ankan explained here, provides diminishing results, beyond around 2.

    Requesting all 8 threads, PTS reports a score of 429 for an RX 580. Manually running the benchmark with 2 threads gets 860 nodes per second, nearly in line with the RX 580 ROCm result Michael got. Finally, specifying one million nodes (as ankan suggested for properly testing the TITAN RTX) brings the score up to 1100.


    • #12
      Originally posted by alcalde View Post

      I thought neither TensorFlow nor PyTorch worked with OpenCL?
      TF has a compiling option for OpenCL, but I have not tested the OpenCL backend and always use CUDA because
      1. I'm not using AMD's GPU for deep learning -- for now.
      2. NVIDIA apparently implemented their OpenCL with CUDA so it takes ~10% overhead...