American company Cerebras Systems, which specializes in designing AI chips and systems, announced a new server called CS-4. It is based on three large chips, each about the size of a dinner plate. The company focuses on accelerating the inference stage, where the model uses what it has learned to produce an answer for the user within chatbots.

According to a report published by Reuters, the server relies on the Nexus architecture and modular units containing WSE-3 Turbo chips. Cerebras says that the design gives it higher speed than traditional systems because it reduces the need to constantly transfer data between a large number of small chips. This process consumes energy and adds time to generating an answer.

Larger size, less data transfer
 

The new chips are made by TSMC, a Taiwanese company that specializes in manufacturing semiconductors, using 5-nanometer technology. Cerebras has redesigned the server to include 50% fewer components, which is expected to simplify its assembly and speed up data center provisioning. The system will be available to customers in the third quarter, with new communication components designed to increase the speed of data transfer between the three chips inside the server.

Nvidia's competition shifts to answer speed
 

Cerebras is betting on the model deployment market, rather than just training it. This market is expanding with the increased use of chatbots and instant services. The company expects to deliver computing systems with a total power of 600 megawatts by the end of 2027. It also aims to increase speed fourfold and processing capacity twentyfold. Cerebras plans to launch a new generation of chips and servers in 2027, in an attempt to gain a share of the market currently dominated by Nvidia.