Cerebras says a better cooling design lets its wafer-scale chip run harder, powering a new rack system it calls the CS-4.
At the center of the CS-4 sits the WSE-3T, a four-trillion-transistor processor that spans an entire wafer. Its 44GB of on-chip SRAM feeds 43.2 PB/s of memory bandwidth, and 2.4 Tbps of off-die connectivity links the racks, while the chip sustains 250 petaFLOPS of AI compute.
Cerebras argues that holding model weights on the chip, rather than fetching them from external memory, makes the CS-4 especially fast at generating results. A few select customers are testing the system now, with broad availability slated for the third quarter.
The launch comes as Cerebras seeks to justify its $5.55B May IPO. Second-quarter results showed a $450M loss on $180M of revenue, though $377M of that red ink traces to IPO-related stock compensation.
