This research introduces LLMPerf, a novel approach that leverages Large Language Models to predict GPU program performance without traditional manual modeling.

• Uses LLMs as performance estimators for OpenCL programs on GPUs
• Achieves significant accuracy improvements over conventional methods
• Demonstrates LLMs' ability to understand complex hardware-software interactions
• Provides a more scalable approach to performance prediction across diverse GPU architectures

For Engineering teams, this breakthrough enables more efficient resource allocation, faster development cycles, and optimized GPU utilization without requiring specialized performance modeling expertise.

LLMPerf: GPU Performance Modeling meets Large Language Models