Cerebras unveils WSE-3 Turbo and CS-4 system, delivering 30x inference speed over GPU

Cerebras, a wafer-scale AI inference accelerator maker, launched its new WSE-3 Turbo chip and CS-4 system on the 18th local time. Compared to the previous CS-3, CS-4 offers 10x token capacity and 2x speed. The WSE-3 Turbo packs 900,000 cores and 44GB of SRAM, with doubled FP16 sparse AI compute, memory bandwidth, on-chip interconnect, and I/O bandwidth. The CS-4 can integrate three WSE-3 Turbo chips, boosting compute density.