Anyone going to repurpose a bc250 crypto GPU for local AI

A community kernel patch by S. Duggan re-enables 16 factory-masked Compute Units on the AMD GFX1013 die of the bc250 crypto GPU, boosting median generation speed by 1.32× and median prefill speed by 1.50× across 11 LLM models in llama.cpp b9265 at 4K context. The unlocked 40-CU configuration raised generation speeds from 104 to 129 tok/s on Granite 4.0-H Tiny, 66 to 87.5 tok/s on GPT-OSS 20B MXFP4, 59.5 to 78.7 tok/s on Qwen3.5 35B-A3B MoE, and 9.6 to 14.9 tok/s on QwQ-32B IQ2_M, with power draw rising from 101 W to 116 W median.

TL;DR: The GFX1013 die physically has 40 Compute Units; the factory firmware masks 16 of them. A community kernel patch by S. Duggan re-enables the remaining CUs. After an independent FP32 sanity check 100M error-free multiply-adds , a controlled paired A/B re-run on 11 models followed. Key results at 4K context, llama.cpp b9265, verified-stock 24 CU → patched 40 CU: - Median generation speed-up: 1.32× all 11 deltas positive, p < 0.01 - Median prefill speed-up: 1.50× prefill gains more because it’s more compute-bound - Granite 4.0-H Tiny hybrid Mamba : 104 → 129 tok/s +24% - GPT-OSS 20B MXFP4: 66 → 87.5 tok/s +32% - Qwen3.5 35B-A3B MoE: 59.5 → 78.7 tok/s +32% - QwQ-32B IQ2 M reasoning dense : 9.6 → 14.9 tok/s +55% The speed-up is larger for prefill than generation, consistent with generation being memory-bandwidth-bound on this hardware adding compute units helps less than reducing weight bytes per token . A roofline measurement confirms: peak streaming bandwidth 357 GB/s, peak FP32 3901 GFLOP/s, ridge point 10.9 FLOP/byte — all LLM decode quantizations sit left of the ridge bandwidth-bound regime . The unlock is bounded by power/cooling: the 40-CU arm ran at 116 W median vs 101 W at 24 CU, holding the clock ~3–4% lower under the oberon self-throttling governor. A better-cooled unit would likely gain more. The long-context picture is strong. The Qwen3.5 35B-A3B MoE runs to 64K filled 16.6 tok/s, n=3 — provided ttm.pages limit is genuinely at the full 16 GiB §3.3 . The dense qwen3.5:9b also runs to 64K 16.4 tok/s and retrieves all 5 needles at 128K; its smaller footprint leaves more headroom alongside the always-on services, so it’s the safer production default — but the MoE is now a genuine long-context option, not short-context-only. And the 40-CU unlock carries this into long context : re-run under the 14.7 GiB working-set guard, all four heavy MoEs climb the full 4K→64K ladder at 40-CU with 2/2 needle retrieval at every tier — the flagship Qwen3.5-35B-A3B at 74→48 tok/s — so the active-parameter speed advantage survives the long-context regime under the unlock, not just at stock 24-CU §B9.2 .