BC-250 Mining Board Trains PyTorch Faster Than a Tesla T4

The ASRock BC-250 mining board with a cut-down PS5 APU now hits 2,337 train images per second in PyTorch using Mesa rusticl and a patched OpenCL backend.

Linux / Steam Deck, PC (Windows) · Hardware

0xKova5 Nov 2026, 08:35· watcher:bc250

The ASRock BC-250, a surplus crypto-mining board built around a cut-down PlayStation 5 APU, can now train PyTorch models on its GPU through OpenCL. Developer mxreyer updated the pytorch-gpu-bc250 repository with simpler setup instructions and new benchmarks for GPU-accelerated PyTorch on the board’s gfx1013 graphics block.

Because AMD’s ROCm stack leaves out gfx1013 support, the workaround relies on Mesa’s rusticl OpenCL driver paired with a patched pt_ocl.so backend from mxreyer/pytorch-dlprim-gfx1013. Bypassing ROCm in favor of plain render nodes means containers only need --device /dev/dri/renderD128 to access the hardware.

Benchmarks run ResNet-9 on CIFAR-10 in fp32 at 40 CUs and 2.0 GHz. The patched setup trains at 2,337 images per second and infers at 7,130. A Google Colab Tesla T4 running the same workload reaches 2,294 and 7,217, so the BC-250 trains slightly faster and infers slightly slower. Test accuracy stays between 0.922 and 0.927 on every platform. Stock pytorch_ocl reaches only 919 training images per second at the same clock, so the patch does most of the work. The 40 CU and 2.0 GHz settings depend on an unlock, which the repo links to. Enabling fp16 convolutions pushes training to 2,992 images per second.

Sources: github.com