Cerebras Launches CS-4, a High-Performance AI Inference Solution

The Cerebras CS-4 is a rack-scale AI inference solution that offers up to 30 times faster inference than GPU systems.

Each wafer in the CS-4 delivers double the speed of its predecessor, the CS-3, while achieving up to 10 times more throughput per watt.

The system reduces interconnect latency to 2 microseconds, enabling over 1,000 tokens per second for models with more than 10 trillion parameters.

CS-4 is built on the new Cerebras Nexus Platform Architecture, featuring a modular design that simplifies manufacturing, deployment, and maintenance.

The Wafer-Scale Backpack integrates the wafer, power conversion, cooling, and control electronics into a compact package, reducing deployment time significantly.

Power delivery is optimized, being only 0.5 millimeters from the processor, which minimizes power loss and allows for higher operating frequencies.

A new programmable I/O subsystem in CS-4 doubles I/O bandwidth and lowers latency, facilitating connections between wafers without the need for a switch.

The Cerebras PowerRack can be pre-installed, allowing compute backpacks to be added quickly, streamlining the deployment process in hyperscale environments.