Cerebras unveils CS-4 rack built around WSE-3 Turbo

Cerebras has introduced the CS-4, a rack-scale AI system powered by its WSE-3 Turbo, or WSE-3T, wafer-scale chip. The processor retains the company’s largest AI processor positioning with 4 trillion transistors, 900,000 cores and 44 GB of SRAM integrated on the wafer, while delivering 125 PFLOPs of AI compute, or 250 PFLOPS with sparsity, per wafer.

The WSE-3T also raises bandwidth to 43.2 PB/s, with an on-chip fabric rated at 53.5 PB/s and off-chip I/O at 2.4 Tb/s. Cerebras says the updated engine cuts latency from 5ms to 2ms, aiming to make fast inference practical on large frontier models rather than relying on smaller systems.

CS-4 uses a new Nexus Platform Architecture organized around compute, power and I/O layers, with wafer-scale modules attached as pluggable backpacks. Each rack includes three WSE-3T (Turbo) chips and is rated for 750 PFLOPs of AI compute, 7.2 Tb/s of I/O bandwidth and 129.6 PB/s of SRAM bandwidth.

Cerebras claims a single CS-4 rack can generate over 4,400 tokens per second in GPT-OSS 120B, and says a GPU rack took 30 secs to produce what CS-4 generated in one second. Shipments are expected to begin this quarter, with large CS-4 clusters designed to support models with over 50 trillion parameters.