Cerebras Introduces CS-4 System for AI Inference, Claiming Up to 30x Faster Than GPUs

Cerebras CS-4 Unveiled for AI Inference

Cerebras has announced the CS-4, a new rack-scale solution aimed at accelerating AI inference. This system incorporates three WSE-3 Turbo wafers, each providing up to twice the speed of the previous generation. The CS-4 is presented as an architecture for frontier AI applications.

Performance Claims and Throughput

The CS-4 is stated to deliver up to 30 times faster inference compared to GPU systems, setting a new record for production inference speed. It also claims to offer up to 10 times more throughput per watt than its predecessor, the CS-3, while generating tokens up to 30 times faster than production GPU systems. This design focuses on both high throughput and interactive performance.

Hyperscale and Frontier AI Capabilities

Designed for hyperscale datacenters, the CS-4 is the first iteration of the Cerebras Nexus Platform Architecture. It features reduced wafer-to-wafer interconnect latency of 2 microseconds, enabling it to process over 1,000 tokens per second on models exceeding 10 trillion parameters, maintaining interactive decode performance at scale.

Modular Design and Power Delivery

The CS-4 utilizes a modular compute backpack design, integrating the wafer, power conversion, liquid cooling, high-speed I/O, and control electronics into a compact package with 50% fewer components. This design aims to simplify manufacturing and reduce deployment time. The system also features high-density power delivery, with power components located 0.5 millimeters from the processor, which is stated to nearly eliminate board-level power loss and enable higher operating frequencies for the WSE-3T.

Next-Generation I/O

A new programmable I/O subsystem is introduced with the CS-4, which doubles I/O bandwidth and reduces latency. This enhancement benefits both aggregated and disaggregated solutions within the system.