1e29 FLOPS would take 200k GB200s 3.2 years and 6.2 TWh, dev notes
mike64_t · x · 2026-11-03
The author mocks claims about training a quadrillion-parameter neural network at 1e29 FLOPS, showing the math: it would require 200k GB200 GPUs running at 100% MFU in FP8 for 3.2 years, consuming at least 6.2 TWh under idealized conditions — arguing such scale is far removed from reality.
More from Fun
- 'Definitely not Mario': AI-generated copyright-dodging game characters go viral — zeeg · 2026-11-03
- Everyone Is Building a VCS Replacement — and It's Turning Into a Following Frenzy — latkins · 2026-11-03
- Sentry CEO David Cramer open-sources Slag, a party platformer built with Claude Code — zeeg · 2026-11-03
- AI launch culture debate: paid engagement is a worse look than an earnest, cleanly executed launch — Jackyhuang · 2026-11-03
- Suffering is like pH: essential at 2/10, horrifying at 8/10 — louisvarge · 2026-11-03
- X users report a sharp surge in bot replies over the past few days — jaivin · 2026-11-03