LUMINA: LLM-Guided GPU Architecture Exploration via Bottleneck Analysis

LUMINA is a new LLM-driven framework for GPU architecture exploration that uses AI to optimize GPU designs for modern AI workloads like LLM inference. The system achieved 17.5x higher efficiency than traditional methods and identified 6 designs superior to NVIDIA's A100 GPU using only 20 exploration steps.

• →LUMINA leverages LLMs to automate GPU design space exploration, reducing the need for manual analysis and human heuristics.
• →The framework identified 6 GPU designs with better performance and area efficiency than NVIDIA's A100 using minimal computational resources.
• →LUMINA achieved 17.5x higher exploration efficiency and 32.9% better design quality compared to traditional machine learning baselines.
• →The system includes a comprehensive DSE Benchmark for evaluating LLM capabilities in architecture optimization tasks.
• →From a design space of 4.7 million possible configurations, LUMINA found optimal solutions in just 20 steps through bottleneck analysis.