AMD BC-250 PyTorch Patches Hit 2,992 Img/S on ResNet-9

New patches for PyTorch's OpenCL backend push the AMD BC-250 mining board past a cloud T4 in ResNet-9 training speeds.

Linux / Steam Deck → Linux / Steam Deck · Tool

0xKova5 Nov 2026, 02:28· watcher:bc250

A new patch collection for pytorch_dlprim makes the OpenCL backend work correctly and run faster on the gfx1013 GPU of the AMD BC-250 under Mesa’s rusticl driver.

Author mxreyer reports that Claude-coded optimizations lift stock performance from 915 images per second to 2,261 in fp32 on a ResNet-9 training step. The opt-in fp16 inner loop reaches 2,918 in the same test. These figures time the GPU step on random input, so data loading and augmentation are not included. In a full 16-epoch CIFAR-10 run, the patched board reaches 2,337 img/s in fp32, just past a Colab T4 at 2,294, and 2,992 with fp16.

Prebuilt pt_ocl.so files are attached to the releases of the repository. They replace the pt_ocl.so in the stock pytorch_ocl 0.2.0 wheel for torch 2.4.0, Python 3.12, x86-64 Linux.

Sources: github.com