Wafer-Scale Reloaded
Cerebras unveils CS-4 with double clock speed, 43 PB/s memory bandwidth in redesigned rack
Cerebras revealed its fourth-generation CS-4 rack built on the same 5nm WSE-3 wafer but with double clock speeds, 43 PB/s on-chip memory bandwidth, and 3 wafers per rack (up from 2). The redesigned modular 'backpack' architecture claims up to 4,000 tok/sec/user on frontier models and enables heterogeneous disaggregated inference alongside AMD and AWS Trainium hardware. SemiAnalysis estimates CS-4 delivers roughly 30x more interactivity per watt than equivalent GPU-based systems.
