A Novel AI-Native GPU Kernel Programming Language for Autonomous Kernel Tuning
A Novel AI-Native GPU Kernel Programming Language for Autonomous Kernel Tuning
This SIGOPS ATC 2026 paper presents an AI-native GPU kernel programming language co-designed with AI tuning, arguing that existing LLM-driven tuners are limited because they only search parameters within human-supplied skeletons and rely on substrates with large token footprints and no compile-time feedback. In a six-DSL comparison using the same agent, H800 GPU, and iteration budget, its AI-tuned kernels averaged 1.23× over cuBLAS, CUTLASS, and cuSPARSELt across 95 production shapes, with trajectory analysis attributing gains to structural discoveries such as warp specialisation, pipeline staging, and TMA. The authors report compile pass@1 of 85.7% (95.6% pass@5), 82.1% pass@1 on dynamic-shape kernels versus Triton's 48%, and 82.3% under moderate-capacity models versus Triton's 65.8%.
Source: Hacker News AI · al.radbox.org