AMD BC-250 Gets Tuned ROCm Training Stack and Driver Patches for gfx1013
New toolkit brings optimized PyTorch and ROCm training to the AMD BC-250 ex-mining board with custom GEMM kernels.
The AMD BC-250, an ex-mining board with a cut-down PlayStation 5 APU known as Cyan Skillfish or gfx1013, now trains language models faster on Linux. Developer bilawalriaz released bc250-rocm-training, a new open-source toolkit that adds ROCm and PyTorch training support to the hardware.
ROCm has no GEMM kernels tuned for gfx1013, so rocBLAS falls back to generic kernels that slow training down. The package patches the amdgpu driver to enable all 40 compute units and fixes the GPU memory mapping problems that stop ROCm on this board. It also adds Tensile GEMM kernels tuned for gfx1013 training shapes. The driver patch targets kernel 7.3.1-1-cachyos, so other kernels need the patch ported first.
Benchmarks show the tuned kernels more than double MiniCPM5-2B training speed, from 193 to 419 tokens per second on a 256x8 batch. SmolLM2-135M full fine-tuning in FP32 goes from 3,345 to 5,685 tokens per second.
To try it out:
- Ensure you run CachyOS with kernel
7.3.1-1-cachyosand ROCm 7.2.4. - Clone the repository and run the preparation script to set up your environment.
- Build the patched kernel module, compile rocBLAS and PyTorch from source, and install the tuned GEMM kernels. The PyTorch build takes several hours and needs 16 GiB of swap.