Product Lineup

Specifications

Overview

Cerebras Systems pioneered the wafer-scale engine (WSE) — instead of dicing a 300mm wafer into hundreds of chips, it uses nearly the entire wafer as a single processor to deliver dramatically higher bandwidth and lower latency than discrete GPU clusters. The WSE-3 Turbo is a faster revision of the third-generation WSE-3 (launched 2024), doubling performance across the board by raising clocks on the same TSMC 5nm, 4-trillion-transistor design.

Each WSE-3 Turbo integrates 900,000 AI-optimized cores with 44 GB of on-wafer SRAM. Rated compute doubles to 250 PFLOPS of sparse FP16, on-wafer memory bandwidth to 43.2 PB/s, mesh fabric bandwidth to 53.5 PB/s, and off-wafer network bandwidth to 300 GB/s. Cerebras has not disclosed exact power, but the CS-4 rack it targets is built to deliver roughly twice the power of the prior system, implying a per-wafer draw near 54 kW — indicating the company tamed its voltage/frequency curve enough to double clocks without a superlinear power penalty.

The WSE-3 Turbo is the compute heart of Cerebras's first true rack-scale system, the CS-4, where three Turbo wafers are linked into a single 750 PFLOPS rack. Cerebras positions the wafer-scale approach as a latency- and bandwidth-advantaged rival to GPU scale-up systems for large-model training and inference.