Cost-driven interest in running large language models (LLMs) on non-mainstream silicon outpaces the maturity of the surrounding software stacks. This paper uses one such platform, the AMD BC-250 (a repurposed cryptocurrency-mining board with a GFX1013 "Cyan Skillfish" accelerated processing unit (APU), 16 GB of unified GDDR6 memory, and a Vulkan-only GPU compute path), as a single-board testbed. Three observations not specific to this board are reported: two benchmark-validity cautions and one controlled empirical result. The cautions are (i) a silent prompt-truncation failure mode in the Ollama runtime that affects filled-context benchmarking unless the API's promptₑvalcount field is verified per request, and (ii) the operational distinction between context allocation and context utilisation, which can produce headline "128K context" numbers backed by a much shorter prompt. The empirical result is (iii) that on hardware lacking matrix accelerators, sparse mixture-of-experts (MoE) generation throughput appears to track the active-parameter count rather than the total. The practical edge over a deployable dense model is modest (a 35 B-total / 3 B-active MoE leads the 14 B dense models by ≈14–25 % at short context) ; the larger ≈6× separation reported here is against a same-size dense control held at the same ≈2. 5-bit quantisation, a configuration chosen to isolate the variable rather than one intended for deployment. Both readings are consistent with a roofline measurement that places decode left of the compute ridge, in a regime bounded by memory bandwidth relative to compute. The evidence rests on a 31-model core cohort (3–35 B parameters), with six further follow-up reference builds added for targeted sub-studies, covering generation speed (statistically validated; within-cell coefficient of variation (CV) ≤ 0. 4 % across the canonical n=3 cells), filled-context scaling with 80 % real-token payloads, output quality across five task types, cold-start latency, and clock and thermal telemetry. A reproducible deployment recipe is documented as the enabling infrastructure for these measurements: a kernel TTM (Translation Table Manager) pagesₗimit adjustment plus an independently FP-validated reuse of a community 40-CU core unlock.
Artur Andrzejczak (Mon,) studied this question.