Cerebras Systems unveiled the CS-4, a rack-scale AI system that links wafer-scale processors across racks without network switches. Cerebras says the system can run up to twice as fast as the CS-3 and, in a GPT-OSS-120B test, deliver up to 30 times the per-user inference speed of GPU-based systems.
The company also says actual throughput varies by model and serving setup. The CS-4 is the first system built on Cerebras’ Nexus rack-scale architecture.
What the Nexus architecture changes inside the rack
Cerebras says Direct Wafer Links can cut wafer-to-wafer latency to as little as two microseconds. The company also says each CS-4 provides up to 7.2 Tbps of system I/O bandwidth.
That internal fabric is proprietary. Third-party accelerators cannot be added to it, and scaling the fabric requires additional Cerebras systems.
How the CS-4 fits into existing data-center networks
The CS-4 supports RoCE v2 over Ethernet so it can connect with existing data-center networks. Cerebras says RoCE v2 allows the system to plug into existing high-speed Ethernet infrastructure.
Outside the Cerebras fabric, the networking picture does not disappear. Manish Rawat, a semiconductor analyst at TechInsights, said the architecture could reduce dependence on traditional switching layers between Cerebras systems but would not eliminate conventional networking requirements.
What this means for AI infrastructure teams
Neil Shah, vice president for research at Counterpoint Research, said eliminating switches between Cerebras systems could reduce infrastructure costs and power consumption. He said direct rack-to-rack links can remove a layer of networking hardware that can consume a large share of an AI cluster’s budget and up to one-third of its electricity.
Danish Faruqui, CEO of Fab Economics, said data ingestion and cluster management would still rely on traditional Ethernet infrastructure, while external systems in disaggregated inference would communicate with the CS-4 over RoCE v2. Cerebras also supports frameworks including PyTorch, but workloads run through the company’s CSoft software platform.
Shah said the Cerebras ecosystem lacks the maturity and breadth of Nvidia’s CUDA environment.