Cerebras unveiled the CS‑4 rack‑scale AI system that connects wafer‑scale processors without network switches, reducing latency and infrastructure costs.
Cerebral Systems has unveiled the CS‑4, a rack‑scale AI platform that eliminates traditional network switches by directly linking its wafer‑scale engines, promising lower latency and reduced infrastructure costs.
Switchless Architecture Overview
The CS‑4 system builds on Cerebras’ wafer‑scale engine (WSE) technology, but unlike previous generations, it uses a proprietary interconnect that ties multiple WSEs together without the need for external Ethernet or InfiniBand switches.
Each rack houses up to eight WSE‑2 chips, each delivering over a trillion operations per second. The direct‑connect fabric provides sub‑microsecond communication between chips, a stark contrast to the microsecond‑scale latency introduced by conventional switch‑based fabrics.
Benefits for AI Workloads
By removing switches, the CS‑4 reduces the number of network hops, which translates to faster model training times, especially for large language models that require frequent synchronization across nodes.
The architecture also cuts power consumption and hardware costs, as data centers no longer need to provision and manage high‑density switch fabrics alongside the AI racks.
Comparison with Traditional Clusters
Conventional AI clusters rely on a hierarchy of switches to connect GPUs or custom ASICs, introducing both latency and points of failure. Cerebras’ switchless design sidesteps these issues, offering a more deterministic performance profile.
- Direct wafer‑scale interconnect eliminates switch latency
- Simplified cabling reduces deployment complexity
- Lower total cost of ownership due to fewer components
Early adopters are expected to evaluate the CS‑4 for workloads such as transformer training, reinforcement learning, and high‑resolution image processing, where inter‑node communication can become a bottleneck.
"The switchless approach could redefine how we think about scaling AI hardware," said a senior engineer at a leading cloud provider.
Cerebras plans to ship the CS‑4 later this year, with pricing details to be announced alongside the product launch.