Cerebras has announced the CS-4, a new AI system that combines three wafer-scale processors in a single rack. The manufacturer focuses primarily on inference and claims higher speeds and better energy efficiency than its previous generation.
The CS-4 is the first system based on Cerebras’ new Nexus architecture. It contains three WSE-3 Turbo processors (WSE-3T), which are manufactured by TSMC using a 5nm process, according to Reuters. Each chip contains four trillion transistors and 900,000 AI-optimized cores, with 44 GB of SRAM directly on the wafer.
According to Cerebras, the processors collectively deliver 750 PFLOPS of AI computing power and 129.6 petabytes per second of memory bandwidth.
Less data traffic between chips
The scale of the processors is central to Cerebras’ approach. In conventional GPU systems, data must be moved between separate chips, which consumes time and energy. Because much of the processing takes place within a single wafer, Cerebras can minimize that data traffic.
In a test using GPT-OSS-120B, the CS-4 reportedly processed more than 4,400 tokens per second per user. Cerebras claims speeds that are up to thirty times faster than GPU-based solutions in certain configurations. The results come from the company itself and vary depending on factors such as model architecture, context length, and configuration.
Compared to the CS-3, Cerebras claims double the speed and up to ten times more throughput per watt.
Direct connection between wafers
The CS-4 offers a total I/O bandwidth of 7.2 terabits per second. With Direct Wafer Links, processors can also communicate directly without a switch. According to Cerebras, latency between wafers drops from five to at least two microseconds.
This makes it possible to combine multiple systems into larger clusters. Cerebras states that the architecture can support models with more than 50 quadrillion parameters. The I/O module also supports RoCE v2 RDMA over Ethernet for connections to other hardware.
The physical design has also been modified. The processors are housed in modular Wafer-Scale Backpacks that integrate power conversion, liquid cooling, and I/O, among other features. According to Cerebras, these modules contain approximately half as many components as their predecessors.
Next generation in 2027
The first CS-4 systems will be delivered in the third quarter. A price has not been announced. According to Reuters, Cerebras plans to release the next generation of the processor and server system as early as 2027.
The company expects to deliver 600 MW of compute capacity by the end of 2027. CEO Andrew Feldman says that by that time, future systems should be four times as fast and offer twenty times the throughput. Last week, Cerebras reported revenue of $180.1 million and an adjusted loss of $6.9 million.