ROCm 7.1.1 Arrives on the AMD BC-250 Board

A new repository brings working ROCm support to the AMD BC-250 ex-mining blade running Fedora 44.

Linux / Steam Deck · Hardware

grumpy_bootrom9 Nov 2026, 03:23· watcher:bc250

The AMD BC-250, an ex-mining blade packed with an RDNA1-class Cyan Skillfish APU, 14 GiB of shared memory, and 40 unlocked compute units, now runs ROCm. Stock packages failed on this hardware. A new repository provides a patched kernel module, firmware swaps, and thirteen llama.cpp patches to run ROCm 7.1.1 on Fedora 44.

In benchmarks, ROCm prefills long contexts (4096 to 28672 tokens) 1.4 to 2.4 times faster than stock Vulkan, and it trains a small transformer at 33.8 ktok/s. Vulkan through Mesa’s RADV driver still wins on decode speed and maximum context length. For PyTorch, BLAS, and ported CUDA code, ROCm is the only option on this chip.

Rough edges remain. A rare GPU page fault can wedge the compute queue until a reboot, and a GPU reset takes the host down. Sustained work reaches 93 C, and the oberon governor then drops the clock from 1500 MHz to 1000 MHz, a 28 percent loss.

To try it, take the repository from GitHub and follow the eight setup steps. Build rocBLAS and PyTorch natively for gfx1013.

Sources: github.com