Get Started
Home
Topics
Search
Library
Research questionHow can LLMs reliably construct analytical hardware-performance models from workload, architecture, and mapping specifications?LLMs can give plausible answers about hardware mappings while still failing to produce reliable analytical models. The difficulty is translating computation, data reuse, storage, and movement into models that correctly capture performance requirements.
AI
Code Generation & Program Synthesis
Evaluation & Benchmarks
Reasoning
Technology
Latest papersRecent research connected to this question, newest first.PerfReasoning: How Well Do LLMs Reason on Hardware Performance?The evidence comes from the PerfReasoning benchmark, which evaluates LLMs on direct performance reasoning and analytical performance-model code generation using workload, architecture, and mapping specifications. Its tasks include comparing mappings and predicting off-chip traffic and buffer requirements; results show a substantial gap between reasoning-based question answering and reliable model construction, with task-specific reinforcement learning improving a small model's mapping-reasoning accuracy while feedback-free self-revision is inconsistent.research paper · Sep 3, 2026
Related questions
How can LLMs synthesize executable machines and improve their designs through physical simulation feedback?How can LLMs reliably retrieve and reason over complex tables in 3GPP specifications, with or without table context?How can LLM-generated GPU kernels remain performant across hardware, inputs, and programming models?How can LLMs generate reliable, adaptive tests that expose one another’s model-specific weaknesses?