Cerebras Unveils CS-4

Cerebras Systems unveiled its fourth-generation wafer-scale system, CS-4. The CS-4 is a system housing three units of the new WSE-3 Turbo processor in a single rack, achieving inference speeds up to 30x faster than GPUs and up to 2x faster than the previous generation CS-3, measured in tokens per second per user (TPS/user). Total token throughput increased by up to 10x compared to CS-3 within the same power budget.

The WSE-3 Turbo maintains the same transistor count and SRAM capacity as the existing WSE-3 but doubles compute performance and bandwidth by modifying the rack design to shorten power delivery distances and increase operating frequency. The CS-4's system memory bandwidth reaches 129.6PB per second, thanks to an architecture that reads weights from on-wafer SRAM.

The new Nexus rack-scale platform separates compute, power, and I/O into independent modules to support flexible scaling. Notably, it is designed to maintain interactivity in large models with 10 trillion parameters by reducing I/O latency to 2 microseconds. Cerebras emphasized that this system can significantly increase the number of inference and verification steps in agent systems, thereby enhancing AI productivity.