Self-Optimizing GPU Kernels for Computational Chemistry: An Autonomous Acceleration Laboratory

<jats:p>Accelerating scientific code on GPUs is central to high performance computing, but finding the best strategy for a given kernel still demands manual experimentation and deep parallel-programming expertise. We present an autonomous laboratory that automates the whole loop: it takes a sequential kernel, proposes optimization hypotheses, emits real CUDA, compiles and runs it on a remote GPU server, verifies it against a gold reference, and iterates. The laboratory supports three ways of writing the kernel, all judged by one external gate: (A) a deterministic search over a hand-built catalog of GPU techniques, with no language model at run time; (B) a language model, local or cloud, that writes each kernel from scratch; and (C) an autonomous coding agent (Codex CLI) that drives the loop by itself. The single external gate is what keeps every number honest: it first checks correctness (the GPU checksum must match the sequential gold to a set tolerance) and only then checks timing (a roofline floor plus a wall-clock guard prove the</jats:p>