Hardware veteran: low-precision gains nearly exhausted, true sparsity is AI's next 10x
blelbach · x · 2026-09-17
Hardware observer blelbach laid out where AI compute scaling goes next:
- CMOS shrinks have delivered power efficiency rather than raw speed since 2015, and that path is nearly done;
- New packaging (stacking, substrates) keeps "chip" scaling alive on the datacenter side, with some room left;
- The largest remaining direction is architectural innovation around information density and specialized compute;
- The low-precision gold mine is ending — FP3/FP2 and bitnets may be the last stops.
In a follow-up he notes today's machines are dense linear algebra engines (a few with block sparsity), and radically different chips supporting true sparsity may be needed to find the next 10x.
Related event: Three Eras of Moore's Law: From Free Lunch to Software Hell(4 posts)→
More from Infra
- Jensen Huang: a 1GW NVIDIA AI factory costs $50-60B but generates ~$50B in annual rental revenue — rohanpaul_ai · 2026-09-17
- Ilya warns neoclouds' weak cybersecurity invites rogue AI agents to hijack compute — Miles_Brundage · 2026-09-17
- AMD's free AI Developer Program: $100 cloud credits, Discord access, hardware raffles — wkmyrhang · 2026-09-17
- After AWS me-central-1 loss, dev jokes about explaining the outage to Codex weekly — andersonbcdefg · 2026-09-17
- Crusoe runs 512 AMD MI355X GPUs at 5.75M tok/s in largest MLPerf inference entry — wkmyrhang · 2026-09-17
- Perovskite could lift solar efficiency ceiling from 30% to 45% — and give the US a shot against China — kyliebytes · 2026-09-17