PyTorch runs on the AMD BC-250 through OpenCL, with pinned downloads now

A GitHub repo shows how to run PyTorch on the BC-250 board via OpenCL, with benchmarks against cloud GPUs.

Linux / Steam Deck · Tool

olya_nro6 Nov 2026, 02:32· watcher:bc250

The AMD BC-250 is a surplus crypto-mining board built around a cut-down PlayStation 5 APU. Its GPU reports as gfx1013, which ROCm does not support. The mxreyer/pytorch-gpu-bc250 repo shows how to run PyTorch 2.4 on it using OpenCL. Mesa’s rusticl driver exposes the GPU to PyTorch as ocl:0, and the out-of-tree pytorch_ocl backend handles the rest. Training compiles on rusticl with a patched pt_ocl.so from pytorch-dlprim-gfx1013.

Recent commits separate two builds: a fixes-only build that runs correctly, and a patched build that runs faster. Both v1.3.0 downloads are pinned by sha256. The README table lists the patched build at 2,337 train img/s on ResNet-9 / CIFAR-10, compared to 2,294 on a Colab T4. One commit notes that the reduce rebuild needs the barrier fix and that libclc does not help, while the README still recommends Mesa’s patched libclc.

How to try it:

  1. Use a Mesa build with rusticl. The container requires only /dev/dri/renderD128, not ROCm’s /dev/kfd.
  2. Follow Stage 2 of SETUP.md for the pt_ocl builds.
  3. Run gpu-check.py to confirm PyTorch uses the GPU and not the CPU.
Sources: github.com