ETH's Hoefler at SC26: right-sizing float precision gives 1.8x speedup; SpaDA hits 260 TFlop/s
thoefler · x · 2026-11-19
- ETH Zurich professor Torsten Hoefler closes SC26 with two talks questioning computing fundamentals, each backed by arXiv papers.
- "What Bloats Your Floats?": challenges the assumption that PDE simulations need FP64 — discretization error already acts as a noise floor masking low-order bits. Proposes an SNR-based criterion for safe precision reduction plus an automated data-centric workflow; strategic mixed-precision on GPUs yields up to 1.8× speedup on weather/climate codes without hurting physical fidelity.
- SpaDA: a programming language for spatial dataflow architectures like Cerebras' Wafer-Scale Engine, with explicit placement, dataflow blocks and async/await semantics; implements stencils, pipelines and linear algebra in 14,009 fewer lines than CSL, hitting 260 TFlop/s across 730,000 PEs on one device.
- The core questions: do we need all those bits, and are we moving them through the machine the right way?
More from Infra
- 'Wet lab is the new neocloud': bio lab capacity emerges as AI's next scarce resource — yawnxyz · 2026-11-19
- Microsoft lobbies Trump admin to halt ban on key China-made AI datacenter component, Reuters reports — PAstynome · 2026-11-19
- Microsoft outage postmortem still dodges the key question: why were servers taken offline? — DavidLinthicum · 2026-11-19
- Tencent's Hy4 Flash goes open-weight, hits 148 tok/s on two workstation GPUs — AIFlow_ML · 2026-11-19
- SpaceX targets 1 GW+ of AI compute per month by end of 2027, Musk says — elonmusk · 2026-11-19
- Musk: 100% of superintelligence compute will eventually be in orbit, Earth rounds to 0% — elonmusk · 2026-11-19