Local MiniMax M3.1 Flash and others: memory-clock and loader tuning gains
Running DeepSeek V4.1 Flash, Qwen3.8, GLM-5.3 and MiniMax M3.1 on two AMD Strix Halo boxes plus an RTX PRO 6000, the author traces most gains to config: GPU memory clock stuck in P1 costs ~4.5% bandwidth and 1.65x slower first tokens, and pipelined expert loading cut load time from 109s to 62s.
Same lesson here, many times over. We run DeepSeek V4.1 Flash, Qwen3.8, GLM-5.3 and MiniMax M3.1 on two AMD Strix Halo boxes plus a single RTX PRO 6000. Most of our wins this week were config/plumbing, not new hardware: • GPU memory clock sitting in P1 (13.4 GHz) instead of P0 (14 GHz) during decode: ~4.5% of bandwidth gone, and the first ~16 tokens ran ~1.65x slower in low power states. • Model weights mapped as uncached memory: every tile re-read the same weights from DRAM. Bench looked great, engine didn’t. • One event query per step on Windows WDDM forced early submission of queued copies: 1–2 ms lost per window. • A serial expert loader read 150 GB one expert at a time: load went 109 s → 62 s by pipelining reads (now disk-bound). And many more...