A new rack-scale AI system promises faster deployment and higher performance by combining wafer-scale computing, modular design, advanced cooling, and high-bandwidth connectivity for demanding workloads.
Cerebras has launched the CS-4 rack, a rack-based system designed to increase the performance of its existing silicon by enabling more power to be delivered to the compute hardware. The company says improved cooling allows the system to operate at higher power levels while maintaining performance.
The CS-4 is built around a rack-scale architecture containing modular compute sleds, with Cerebras positioning the design as simpler to build and faster to deploy. The system separates power, compute and input/output functions, while its front-mounted power delivery is intended to simplify rack integration.
At the heart of the system is Cerebras’ Wafer Scale Engine 3T (WSE-3T). According to the specifications shown in the article, each chip delivers 250 petaflops of AI compute, provides 44GB of SRAM, supports 43.2PB/s of memory bandwidth and offers 2.4Tbps of off-die connectivity.
Cerebras argues that fast result generation is an important competitive advantage for AI workloads. Its wafer-scale approach keeps the values underpinning an AI model on the chip rather than repeatedly moving them between the processor and external memory. The company says this can accelerate the delivery of compute results.
The rack is expected to become more widely available during the third quarter, with a small number of customers already considering the CS-4. The design is aimed at demanding AI applications where computing performance, memory bandwidth and deployment speed are important.
The launch also comes as Cerebras continues to face financial pressure. The article reports that the company recorded a $450 million loss on $180 million of sales in the second quarter. Much of that loss was attributed to stock-based compensation connected with its initial public offering, which the article says would have reduced the loss to $73 million without that expense.
