Gemini V4.1 pretraining estimated at ~5e24 FLOPs, two weeks on 4K B300s
teortaxesTex · x · 2026-09-10
Citing an Astra estimate, teortaxesTex puts Gemini V4.1's pretraining compute on the order of 5e24 FLOPs, somewhat above V3 — roughly 3.5M H100-hours, or two weeks on 4,000 B300s. For a 64B-active, 3-4T-total model that's about two months of training. The quoted post speculates a V4.1 Ultra is coming, arguing Google wouldn't emphasize Flash as the smallest tier unless bigger versions exist.
Related event: Estimate: V4.1 Pretraining Took ~5e24 FLOPs(2 posts)→
More from Infra
- Meetas: fully local meeting assistant grounds every claim in transcript evidence, runs a 27B Q4 model offline — Odd-Name-1556 · 2026-09-10
- NVIDIA's BioNeMo Inference Runtime hits public beta, boosting Boltz-2 folding throughput 2.9x — AllThingsApx · 2026-09-10
- US PCB production fell from ~40% to ~4%: an interactive atlas traces one board's supply chain — AnneliesGamble · 2026-09-10
- GPU indices bullish, token indices bearish: compute appreciates as intelligence gets commoditized — sudoraohacker · 2026-09-10
- Understanding FlashAttention: A Handbook Tracing FA1 to FA4 and Why HBM Traffic, Not FLOPs, Is the Bottleneck — techNmak · 2026-09-10
- Stealth startup Kepler Computing raises $468M for EUV-free HBM alternative using 3D ferroelectric stacking — Promptmethus · 2026-09-10