modded-nanogpt cuts another 2.17 seconds in a 76.0-second speedrun
kellerjordan0 · x · 2026-07-28
A GitHub pull request for modded-nanogpt claims a new world record by cutting the run to 76.0 seconds, after an additional 2.17-second improvement. The change removes sparse attention gates and improves XSA lowering, with the maintainer noting that equivalent math plus better execution already pushes the benchmark to 17.33 seconds on a single H100. The thread frames this as an incremental but meaningful speedrun in optimized training code.
More from Infra
- SK hynix pre-earnings note warns headline numbers may hide stronger memory pricing — tengyanAI · 2026-07-28
- Bloom Energy earnings are now a test of AI power demand quality, not just EPS — tengyanAI · 2026-07-28
- Rescript turns browser-based transcript editing into a free local Descript alternative — sven_ai · 2026-07-28
- Distributed Systems opens graphsub, a Rust-backed substrate with MCP, Jepsen tests and benchmarks — arthurcolle · 2026-07-28
- Distributed Systems open-sources DSCO, a C-written agentic LLM CLI with MCP and swarms — arthurcolle · 2026-07-28
- A meme hides another NanoGPT micro-optimization in the XSA update path — burny_tech · 2026-07-28