AMD BC-250 Stack Reaches Release 10.2 With Native ROCm and PyTorch Training Support
Release 10.2 for the AMD BC-250 brings native ROCm and PyTorch training support to the gfx1013 ex-mining board.
The AMD BC-250 is an ex-mining board housing a cut-down PlayStation 5 APU (Cyan Skillfish, gfx1013, RDNA1). Release 10.2-gfx1013-20261111-138a66f52 for the bc250-stack repository wraps up its ten-package roadmap, delivering native ROCm and PyTorch support.
The drop includes Python 3.12 consumer wheels for torchvision, torchaudio, torchcodec, and W8 quantization via bc250_quant, plus a full ONNX stack and the patched bc250-torch source. Recent commits clear PLAN2-07 FP16 PEFT qualification and admit UNet2D, RNN, LSTM, and GRU models.
Hobbyists must run on gfx1013 only with SDMA disabled, no MFMA/WMMA matrix cores, W8 quantization limits, and a 4-job build cap on ~15 GiB RAM per board. The full 18 GB ROCm prefix is built from source using offline recipes in the repository.