Cerebras CS-4, Rack-Scale AI Inference System

Cerebras announced its new rack-scale AI inference system, CS-4, stating that it delivers up to 30x faster inference speeds and improved economics compared to existing GPU systems. The system is designed for frontier AI architectures and features three WSE-3 Turbo wafers, which are 2x faster than the previous generation.

Based on the Nexus rack-scale platform architecture, CS-4 simplifies manufacturing, deployment, and maintenance through the modularization of compute, power, and I/O. Key technical features include:

• Modular Compute Backpack: Integrates the wafer, power conversion, direct liquid cooling, and high-speed I/O into a single compact 3D package, reducing part count by 50% and cutting deployment time from days to hours.
• High-Density Power Delivery: Delivers power just 0.5mm from the processor to minimize board-level power loss, supplying 2x the power to the WSE-3T to achieve higher operating frequencies.
• Next-Generation Wafer I/O: Introduces a programmable I/O subsystem that doubles bandwidth and reduces latency. It achieves low latency of 2 microseconds for inter-wafer connections without switches, enabling the generation of over 1,000 tokens per second for models with more than 10 trillion parameters.

Additionally, the system maximizes deployment efficiency in hyperscale data centers by adopting an approach where stable power, cooling, and network layers (PowerRack) are installed first, followed by the integration of compute modules.