2026-11-19 · America/Los_Angeles · 论文 · #17
Mechanistic Critics for Sample-Efficient NPU Design-Space Exploration
NPU design-space exploration for LLM inference mainly relies on an expensive simulator that often returns only scalar feedback such as latency and power. Such feedback can rank evaluated configurations, but gives little information for deciding why a configuration is slow or which architecture parameter should be changed next. This limitation is especially severe because prefill and decode workloads can be limited by different resources, including compute throughput and HBM bandwidth. We propose CriticNPU, a sample-efficient DSE framework built around an online-calibrated mechanistic critic…
热度 38.5 / 100;排名与评分保留该期记录。