Answers are grounded in this paper and cite their sources.
Single-board study identifies correctness and liveness problems affecting GPU compute queues, suggesting targeted fixes.
The AMD BC-250 (gfx1013 "Cyan Skillfish" APU) is widely held to have a broken GPU compute queue, leaving ROCm/HIP unusable. This single-board study separates two problems under that one label. A correctness problem, where large kernels silently return wrong results, behaves like a driver regression: a one-line change to the address-translation flush path reliably removes an associated freeze and stall, with no wrong results afterwards on a small sample. A liveness problem, where the queue wedges under load, is traced by a per-wave hardware capture to a shader memory-access violation that halts the dispatch, the reported command-processor timeout being only a later symptom. The same kernel run through the graphics queue stays correct and fast, placing both problems in the compute-queue path rather than the shader hardware. The root cause is not established; all results are from one board.
No takes yet. Share an insight, caveat, or question.
Artur Andrzejczak (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: