BC-250 Stack Delivers Native ROCm and PyTorch for AMD's Cut-Down PS5 APU
The Naster17/bc250-stack project releases version 10.2-gfx1013, bringing native ROCm and PyTorch training support to the AMD BC-250.
The Naster17/bc250-stack project released version 10.2-gfx1013-20261111-138a66f52, turning the AMD BC-250 into a machine learning training device.
The AMD BC-250 is an ex-mining board featuring a cut-down PlayStation 5 APU codenamed Cyan Skillfish, using the gfx1013 architecture. Because upstream ROCm and PyTorch do not target this RDNA1 variant-missing matrix cores like MFMA or WMMA, with SDMA disabled-running local training workloads requires custom patches across the software stack.
The release provides consumer wheels for Python 3.12, a patched PyTorch source tree, an ONNX stack, and a reconstruction record across ten work packages. The umbrella repository coordinates component forks handling kernel quirks in bc250-linux, library repairs in bc250-rocm-libraries, and compiler fixes in bc250-triton.
How to Try It
- Review the requirements and boundaries in the release notes, noting that the board operates under
gfx1013:xnack-with a build cap of four jobs. - Download the pre-built wheels for Python 3.12 or use the offline build recipes in the
recipes/directory to construct the ROCm prefix from source. - Follow the installation steps outlined in the building documentation.