HKUST proposes AccelEval, a benchmark for whether LLMs can automatically accelerate CPU programs into faster end-to-end GPU implementations, accepted at NeurIPS 2026

A team at The Hong Kong University of Science and Technology has proposed AccelEval, a benchmark accepted at NeurIPS 2026 for testing whether large language models can turn correct CPU reference programs into end-to-end faster GPU implementations. It spans 42 tasks across high-performance computing, scientific simulation, graph algorithms, spatiotemporal algorithms, financial computing, and operations research, each at three input scales, and catalogs 43 categories of CUDA optimization strategies for transfer analysis.

In one-pass generation tests on NVIDIA H200, with RTX 4090 cross-hardware checks across eight models, Gemini 3.1 Pro reached a 71.2x geometric mean speedup on medium-scale tasks that passed correctness validation. Code is open source.

Original sources (Chinese)