StringZillas CUDA achieves 32 GB/s multi-pattern matching on single GPU
srchvrs · x · 2026-08-18
Ash Vardanian released a CUDA extension for StringZillas, implementing the Aho-Corasick multi-pattern matching algorithm at 32 GB/s on a single GPU. The implementation can be installed via pip install stringzillas-cuda>=5.2. While noting no major algorithmic breakthroughs, the author claims it may be the fastest matcher available. The related GitHub repo StringWars benchmarks various high-performance string processing libraries.
More from Infra
- Rust Inference Engine Core: 3k Lines, Minimal Tech Debt, Fast Startup — charles_irl · 2026-08-18
- Fal seeks 128 nodes of AMD MI350/355 compute — isidentical · 2026-08-18
- US labs spend 10x more on compute than China, but China builds faster and has easier energy access — Jsevillamol · 2026-08-18
- Tidebroker: Secure Credential Brokerage for AI Agents — steipete · 2026-08-18
- AI Token usage up 1139%, costs down 55% amid caching shift — AccBalanced · 2026-08-18
- US power shortage to leave 43 GW of AI chips idle by 2030 — AccBalanced · 2026-08-18