Zero to 50 ExaFLOPS in Under a Year: Lessons from the Trenches
summary
Cerebras scales an inference service to 50 exaflops in under a year by deploying wafer-scale engines with 900,000 cores and 125 petaFLOPS per chip across liquid-cooled clusters, serving customers like Meta and Mistral. Real fixes are system-level: implementing HTTP2 and origin distribution plus API gateway optimizations halves time‑to‑first‑token, redundancy and multi‑region replicas handle outages, and defenses such as WAF and phone/payment identity plus per‑minute observability dashboards and time‑series stores stop bot floods and pinpoint failures.
AI-written summary · about
chapters (9)
- 0:0050 exaflops in under a year
- 1:24Cerebras wafer-scale engine
- 3:30Adoption growth at scale
- 4:33All components will fail
- 6:45Optimize system not model
- 8:55Evolving request handling
- 10:39Defending against bots
- 12:49Observability never done
- 15:00Operational excellence culture
related talks
34mThe Future of Computing: AI technologies in the exascale era
04m4 real-world AI deployments at enterprise scale
23mKeynote
58mThe Next Decade of Software Development
35mAI innovation begins with trust
23mThe Race to the Bottom: Low Latency in the age of the Transformer
05mFrom Transactions to Transformers: Hardware Built for...
41mBig Data Serving: The Last Frontier. Processing and Inference at Scale in Real Time