{"model":"deepseek-4.1-flash","title":"c1 bottleneck diagnosed (measurement)","status":"positive","summary":"nvidia-smi dmon during sustained PP6 c1 decode: SM ~100% but memory-controller only 4-5% -\u003e c1 is SM-occupancy/latency-bound, not DRAM- or power-bound.","occurred_at":"2026-09-12T08:29:00Z","found":["SM ~100%, mem-ctrl 4-5% at c1 -\u003e kernel occupancy is the wall, not bandwidth"]}