NanoGPT trained in 39.9 seconds, a new world record cutting 46% of training time
cephaloform · x · 2026-09-29
Deven Pietrzak (MIT Math, Hyperstition lab) set a new NanoGPT training record: 39.9 seconds (0.665 minutes), claiming the percentage reduction in training time beats the previous 45 world records combined. The PR (#360) to Keller Jordan's modded-nanogpt was merged after 5 days of work and independently reproduced by four separate sources.
Key techniques:
- ANVIL2 optimization
- Sampled-softmax
- A new embedding table design
- Full-stack fp8 mixed precision
This cut 34 seconds (46%) of training time on the same hardware versus the previous record.
Related event: NanoGPT speedrun record slashed to 39.9 seconds(4 posts)→
More from Infra
- India's Sarvam launches Token Factory, hosting top open models on local infra — HarveenChadha · 2026-11-16
- Gemini 3.5 Flash now runs on air-gapped Google Distributed Cloud, offline up to 12 months — shashib · 2026-11-16
- Gemma.c: full Gemma 1/2/3 multimodal inference in a single C file, ~72 tok/s on CPU at INT8 — This-Purchase-3325 · 2026-11-16
- BIS Issues Interim Final Rule Addressing Remote Access to Advanced AI Chips — apples_jimmy · 2026-11-16
- Jailbroken PS4 runs a 9B open-source model at up to 43 tok/s by talking directly to its 2013-era GPU — AIFlow_ML · 2026-11-16
- Mobile AI Factory: agents auto-build on-device SLMs with 19x less GPU time — pmttyji · 2026-11-16