Skip to content

SC25 Talks at the UMD Booth

SC25 Talks at the UMD Booth

Can LLMs Explain GPU Kernel Performance?

The profiles generated by the Nsight Compute profiling tool are rich in detail but are often challenging to interpret. To achieve maximum performance on a given device, CUDA kernel developers need to spend significant time reviewing and comparing profiles in the tool's graphical interface to understand kernel performance bottlenecks. Large Language Models (LLMs) have shown promise in understanding complex data and generating natural language explanations. In this talk, we propose an agentic LLM framework for interpreting and comparing Nsight Compute profiles to derive useful natural language explanations of kernel performance bottlenecks and suggestions for optimization. We evaluate the framework on a range of CUDA kernels of varying complexity with known performance bottlenecks. We further show that the generated explanations improve the quality of code optimization efforts carried out by LLMs in downstream tasks when provided as context.

Speaker Bio - Joshua H. Davis

Joshua H. Davis

Joshua H. Davis is a fifth-year Ph.D. Candidate and NSF Graduate Fellow at the University of Maryland advised by Prof. Abhinav Bhatele in the Parallel Software and Systems Group. His research focuses on resolving problems related to performance portability across GPU architectures, particularly through the creation of Large Language Model-based tools for performance analysis and code translation. His work on evaluating the performance portability of popular GPU programming models received the Best Research Poster award at SC'23. He has also worked on accelerating high-fidelity CFD applications on GPUs, evaluating LLM capabilities in generating parallel code, evaluating OpenMP GPU offloading capabilities in compilers, and automatic formal verification of concurrent data structures using model checking. In Fall 2026 he will be starting a full-time position at NVIDIA Corporation.