Group Meeting 5

Abstract

In this group meeting, Bowen Cui presented recent advances in AI-powered GPU kernel generation and optimization. He organized the discussion into three main parts: Triton-based approaches (e.g., TritonForge, TritonRL, AutoTriton, TritonGym, GEAK, KernelEvolve), CUDA-based systems (e.g., CUDAForge, CUDA-LLM, robust-kbench, KernelBlaster, Kevin), and broader research insights (e.g., TileLang, PlsemanticsBench, and studies on LLM-driven code quality). The talk emphasized agentic workflows, profiling-guided optimization, reinforcement learning, and the limitations of existing benchmark-centric evaluations, highlighting the need for scalable, hardware-aware, and deployment-integrated optimization frameworks.

Bowen Cui

2024 Fall –

Bowen Cui is a second-year Ph.D. student at George Mason University advised by Prof. Keren Zhou. His research interests focus on LLM-based code agents, benchmarking, and performance optimization.