ETH's Hoefler at SC26: Networks Matter More Than GPUs for AI Supercomputers
thoefler · x · 2026-11-17
Torsten Hoefler (ETH Zurich + Microsoft) previews three SC26 sessions on AI supercomputers, arguing networks matter even more than accelerators. His slides ("Beyond the GPU") cover: a CACM-accepted paper with Dongarra & Matsuoka showing matrix engines + HBM give CPUs GPU-like numerical capability (matched GPU performance at 1.3x energy efficiency and 2.2x cost-efficiency on A64FX, though hardware utilization was <10%); and the SDLA abstraction — program data movement, not just math — pipeline MOVE/COMPUTE/COMMUNICATE via control programs, keep hot data local and overlap the rest. GPUs remain ahead for frontier-scale training, for now.
More from Infra
- NVIDIA's cuPotential speeds molecular dynamics over 3x on Blackwell, eclipsing a decade of gains — AllThingsApx · 2026-11-17
- Resend latency drops across the board: p50, p90 and p99 all trending down — philnash · 2026-11-17
- Local LLM Speed Benchmarks on RTX 4080+3080: Qwen Flash-Next Prefills Faster on One GPU Than Two — Icaruswept · 2026-11-17
- Topos Bio teams with NVIDIA cuPotential to boost molecular dynamics throughput up to 3.6x — AllThingsApx · 2026-11-17
- Infrathon Benchmark Tests LLM Agents on Real-World AI Infra Optimization; Claude Opus 5.5 Leads — RulinShao · 2026-11-17
- PyTorch Day Japan 2026 set for Dec 10 in Tokyo: edge deployment and kernels — PyTorch · 2026-11-17