PerfReasoning: How Well Do LLMs Reason on Hardware Performance?

Dan Zhao

Karthikeyan Sankaralingam

Christos Kozyrakis Stanford

Qijing Huang

Preprint, 2026


Abstract

PerfReasoning evaluates language models on hardware performance reasoning and analytical performance-model generation. Its tasks test computation, data reuse, storage, and movement, revealing a gap between reasoning answers and reliable model construction.