New framework uses LLMs to learn optimal GPU profiling strategies

Researchers have developed AutoProfiler, a novel framework designed to optimize GPU kernel performance by intelligently selecting which hardware information to collect. Unlike previous methods that relied on fixed profiling strategies, AutoProfiler uses a feedback-driven approach with a large language model to learn a dynamic profiling tree. This tree is iteratively refined, allowing for more efficient and effective data gathering, which in turn leads to better kernel optimization, especially under constrained profiling budgets. The findings suggest a shift towards more adaptive profiling reasoning for improved GPU performance. AI

IMPACT Introduces a more efficient method for GPU optimization by leveraging LLMs, potentially improving performance in AI hardware and software development.

RANK_REASON Academic paper detailing a new method for GPU kernel optimization. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• AutoProfiler
• graphics processing unit
• Hugging Face
• KernelBench
• large language models

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →