Metraipolis › archive › dc_8cd8eea0a24f
How Much of a Real Workload Can LLM-Generated GPU Kernels Actually Reach? — arXiv:2609.21058
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2609.21058
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-10T01:07:48.171Z
- SHA-256 of the archived text
f24924230922720cac631aba7cd0fe4c2cef3b61572b67d2fe6b4909bb359253- Stable URL of this page
https://metraipolis.com/documento/dc_8cd8eea0a24f
How to cite
How Much of a Real Workload Can LLM-Generated GPU Kernels Actually Reach? — arXiv:2609.21058 — arXiv (en), https://arxiv.org/abs/2609.21058 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_8cd8eea0a24f (Metraipolis archive dc_8cd8eea0a24f, sha256 f24924230922720cac631aba7cd0fe4c2cef3b61572b67d2fe6b4909bb359253)
BibTeX
@misc{dc8cd8eea0a24f,
title = {How Much of a Real Workload Can LLM-Generated GPU Kernels Actually Reach? — arXiv:2609.21058},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2609.21058},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_8cd8eea0a24f; sha256 f24924230922720cac631aba7cd0fe4c2cef3b61572b67d2fe6b4909bb359253},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.21058. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Language models can now write GPU kernels that outperform PyTorch. We evaluate five model configurations on KernelBench level 1 and find that a frontier model produces correct kernels for 91.1% of problems and independently verified speedups on 22 of 56, including three convolutions, with a median of 1.235x. Open-weights models are far behind: the best reaches 30.4% correct with three verified speedups and solves zero convolutions. We then ask a question the literature does not: what fraction of a real model's wall clock do such kernels govern? Profiling seven workloads across three domains, we find the addressable fraction ranges from 8.9% to 58.2%. On transformers, 80-86% of runtime is spent in cuBLAS GEMM and FlashAttention, bounding realistic end-to-end improvement at roughly 1%, and the fraction shrinks with model scale. On recommenders it is 58.2%, concentrated in a single embedding kernel. We introduce DLRM-Bench, 12 recommender kernel problems in KernelBench format, and measure a 41.7% win rate at a 1.552x median there, projecting 8.63% end-to-end. Separately, we show that KernelBench's correctness check ( this http URL with an absolute tolerance) is satisfied by a tensor of zeros on 4 of 60 level-1 problems. Two kernels in our own results exploited this before we detected them, including one scored at 283x that wrote 0.3% of its output buffer. We propose scale-invariant replacements and release all 879 evaluations. Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-21 in cs.AI). Categorie arXiv: Distributed, Parallel, and Cluster Computing (cs.DC) ; Artificial Intelligence (cs.AI). Pagina del lavoro: https://arxiv.org/abs/2609.21058 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito). Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina. Letto il 2026-10-10T00:23:42.503Z dalla pagina https://arxiv.org/catchup/cs.AI/2026-09-21?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).