Epoch data: leading chips get ~40% more efficient yearly, but AI power draw keeps climbing
teortaxesTex · x · 2026-10-30
teortaxesTex highlights the "Energy and Chips" dossier from Atlas of the Present: Epoch's hardware data shows leading chips 40% more energy-efficient per year, so compute sold growing 2.5x/yr still means fast-rising power draw. A Texas A&M paper shows GPU load swings cut conversion efficiency (89.45% vs 90.96% rated under constant load), and a Stanford/Berkeley study on 220 GDPval tasks finds GLM 5.2 delivers the highest net value ($72.27/task) at 0.430 MJ, while 8 H100s are only 22% faster than 1 but use 3.24x the energy. Author quips that recursive self-improvement toward computronium hasn't arrived yet.
More from Infra
- MediaTek's AI ASIC revenue hits $2.5B in Q4, beats guidance; 448G SerDes tapes out — zephyr_z9 · 2026-10-30
- Gartner: 80% of independent neoclouds will be acquired or fold by 2030 — jeremyakahn · 2026-10-30
- Micron stock drops 6% as Intel sees no end to memory shortage — JOBhakdi · 2026-10-30
- 3 PCs, 120+ threads, 36 coding agents: he says local setups need CPUs, not GPUs — kevinnbass · 2026-10-30
- Open-source Ivory adds an MCP server so AI agents can diagnose database clusters without SSH — aelsergeev · 2026-10-30
- Jensen Huang calls Nvidia's Megaspeed deals "normal" after exec's spying investigation report — pstAsiatech · 2026-10-30