Can't train MoE routers to dodge the bandwidth wall: locality losses fail a 1% perplexity gate
A pre-registered study serving Qwen3-235B on one 8GB GPU finds decode is memory-bandwidth-bound (0.44 tok/s), not compute-bound. Auxiliary locality router losses cut cache misses up to 60% but every config blows past a 1% perplexity budget, confirming miss reduction is tightly coupled to quality. Training-free cache-aware rerouting stacks with trained locality for ~80% miss reduction at lower cost.