Cerebras CS-4 Claims 900,000 Cores, 4 Trillion Transistors and 43.2 PB/s of Memory Bandwidth, Shipping This Quarter

Cerebras / Hacker News·high signalXBlueskyLinkedInCopy link

Cerebras launched the CS-4, built on a WSE-3 Turbo wafer with 900,000 AI cores, 4 trillion transistors, 250 PFLOPS and 43.2 petabytes per second of memory bandwidth, three processors per Nexus rack-scale system. The company claims up to 30x faster inference than production GPU systems, 10x more throughput per watt than CS-3, and more than 1,000 tokens per second on models above 10 trillion parameters, with wafer-to-wafer interconnect latency down to 2 microseconds. First shipments begin this quarter; no pricing was published. The HN thread drew 458 points and 269 comments, with the usual argument about whether Cerebras benchmarks survive contact with real serving workloads.