MoE-on-NPU M0: per-expert GEMM viability (VIABLE) + IRQ-affinity tuning lever - New Exp-2026-06-26-MoE-Expert-NPU-Viability: M0 verdict VIABLE (NPU beats CPU at every Qwen3-A3B expert shape + every M in 8..128, direct int4->int8 inflate collapses materialization 16-30x; bit-exact). Index entry in Experiment-Log. - CPU-Big-Core-Contention: IRQ affinity vs thread-pin distinction (submit threads -> big, HW-IRQ handlers -> little); IOVA anti-fragmentation already-covered note. - Optimization-Roadmap: MoE-on-NPU saturation milestone thread M0..M3 + progress log.
roadmap cleanup: reframe int4 as 'not yet viable' (not dead-end) + what would flip it; flip NEON-repack flag to landed; add CPU-Big-Core-Contention + Quantization-Paths-and-Baselines pages + Home links