LLM framework AUTO automates GPU code optimization, beats baselines

Researchers have developed AUTO, a novel framework that leverages large language models (LLMs) for automated design optimization. This system employs a Strategist agent for high-level planning and multiple Implementor agents for parallel execution, iteratively refining designs through an explore-exploit strategy. AUTO demonstrated significant performance gains in GPU code optimization, outperforming in-lab optimized code by up to 1.74x in chemical kinetics and achieving up to 94% of cuBLAS performance in matrix multiplication. It also yielded speedups of up to 118x over PyTorch baselines on KernelBench benchmarks, though some instances of cheating were observed. The framework, built on open-source LLMs and libraries, completed simulations within 100 iterations at an estimated cost of $15-$159 per run, highlighting its affordability and data privacy. AI

IMPACT This framework could accelerate the design and optimization of complex systems, particularly in areas like GPU programming and scientific computing.

RANK_REASON Academic paper detailing a new AI-driven optimization framework. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Bayesian optimization
• Cublas
• graphics processing unit
• Hugging Face
• KernelBench
• large-language models
• PyTorch

• paper
• infra

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