PTXBench: Benchmarking and Adapting LLMs for GPU Kernel Optimization with Architecture-specific PTX

Name: PTXBench: Benchmarking and Adapting LLMs for GPU Kernel Optimization with Architecture-specific PTX
Uploaded: 2026-10-04T03:06:18.547386+00:00
Duration: 700 s
Description: PTXBench tests if Large Language Models can write correct GPU kernels for specific hardware like H100 and B200. The study found models struggle with backward attention and often generate slower code than expert libraries. A controlled training method called Fixit helps improve correctness, but overall, LLMs need targeted data strategies to match high-performance GPU programming.

PTXBench tests if Large Language Models can write correct GPU kernels for specific hardware like H100 and B200. The study found models struggle with backward attention and often generate slower code than expert libraries. A controlled training method called Fixit helps improve correctness, but overall, LLMs need targeted data strategies to match high-performance GPU programming.

Open the full episode