Cerebras CS-4: 30x Faster Than GPUs (www.cerebras.ai)

Cerebras has unveiled the CS-4, its fourth-generation AI accelerator, boasting a remarkable 30x increase in inference speed compared to traditional GPU systems. This new system is powered by three Wafer Scale Engine 3 Turbo processors and is designed to optimize key elements—compute, power, cooling, and I/O—simultaneously, ensuring a cohesive advancement in AI infrastructure. With capabilities that can serve over 1,000 tokens per second for models with over 10 trillion parameters, the CS-4 aims to provide both high interactivity and throughput, appealing to both developers and data center operators. The significance of the CS-4 lies in its groundbreaking architecture, which enables modularity and flexibility. The Nexus rack-scale platform reimagines system structure with innovative modules for compute, power, and I/O, drastically reducing the complexity of manufacturing and deployment. Additionally, it supports disaggregated inference, allowing for efficient prefill processes paired with its ultra-fast decoding capabilities. With enhanced power delivery and a new low-latency I/O subsystem, CS-4 sets a new benchmark for performance in AI/ML, promising to reshape how large-scale models are deployed and executed in data centers.