Cerebras's Next Generation CS-4: Fast Just Got Faster (newsletter.semianalysis.com)

Cerebras has announced the CS-4, a significant upgrade to its previous generation of AI hardware, set to be detailed further at the upcoming Hot Chips conference. This fourth-generation rack features the same 5nm wafer-scale engine (WSE-3) but achieves double the performance of the CS-3 by increasing power consumption, clock frequency, and rack-scale density. The CS-4 will yield double the tokens per second per user, making it an attractive option for customers looking to maximize revenue with minimal additional investment. Its modular architecture allows for quicker deployment and enhances manufacturability, while the introduction of a new I/O module supports a flexible and disaggregated inference infrastructure, addressing memory constraints more effectively. Key enhancements include the doubling of memory bandwidth and off-wafer I/O capability, reaching 2.4Tb/s compared to CS-3. The unique "Backpack Rack" design streamlines assembly and cooling, allowing three WSEs per rack instead of two, resulting in a total on-chip memory bandwidth of 43PB/s. Nevertheless, a notable limitation remains: the SRAM capacity per wafer remains unchanged at 44GB, restricting memory scalability. By emphasizing disaggregated inference setups, particularly with potential collaborations with AWS and AMD, Cerebras positions the CS-4 not just as a performance upgrade but as a strategic shift towards heterogeneous architectures, aiming to optimize workflows in versatile use cases while addressing the growing demand for high interactivity in AI applications.