30x Token Speed Challenges the GPU Throne! Fabai Praises Cerebras (CBRS.US) CS-4 for Opening the Era of "Wafer-Level + Decoupled Inference"

After the Supernova Day event held by the AI chip company Cerebras Systems (CBRS.US) on Tuesday, BNP Paribas expressed that the partnerships and related announcements made by the company left a strong impression. BNP Paribas analyst Karl Ackerman wrote in a report to clients: Similar to NVIDIA Corporations (NVDA.US) LPU, the new partnerships Cerebras has formed with OpenAI, Amazon.com, Inc. (AMZN.US) AWS, and AMD (AMZ.US) underscore the growing market demand for improvements in latency-sensitive decoding stages during the inference process. The analyst further stated that in terms of rapid inference, the newly announced SRAM-based CS-4 Wafer-Scale Engine provides an astonishing 43.2 PB/s on-chip memory bandwidth and is equipped with a new wafer I/O interface that utilizes RDMA. Cerebras proudly claims that the CS-4 system can generate tokens 30 times faster than current GPU systems in production environments. Additionally, the analyst noted that Cerebras's product roadmap, which aims to double performance annually and increase throughput by 20 times by 2027, is admirable. The analyst mentioned, Cerebras is focusing on decoupled inference, positioning its wafer-scale engine as a high-speed decoding engine to enable rapid token generation, in collaboration with compute-intensive systems like AWS Trainium or AMD Helios. Cerebras expects its decoupled inference system to be up to 10 times faster than GPUs and to achieve a fivefold performance improvement compared to configurations using only WSE. Finally, the analyst indicated that Arista Networks (ANET.US) has become a key network partner for Cerebras and anticipates that, over time, this partnership will help Arista expand into new customer application scenarios. It is reported that Cerebras officially launched its next-generation rack-scale AI system, Cerebras CS-4, at the Supernova Day event on Tuesday, claiming it to be the industry's fastest AI accelerator. This system is targeted at large model inference and directly competes with NVIDIA Corporation's GPU server products. Moreover, during the event, Cerebras announced partnerships with companies such as OpenAI, AMD, Arista Networks, AWS, Figma, Cognition, and CrowdStrike. CS-4 is the first product of Cerebras's new generation Nexus rack-scale platform architecture, powered by three brand new Wafer Scale Engine 3 Turbo (WSE-3T) wafer-scale processors. The WSE-3T is the largest AI processor to date, with each chip integrating 40 trillion transistors and 900,000 AI-optimized cores, covering a silicon area of 46,225 square millimeters, and built-in 44GB SRAM. The CS-4 system offers 750 PFLOPs of AI computing power, 129.6 PB of memory bandwidth per second, and an I/O throughput of 7.2 Tbps per second. The AI computing power of a single WSE-3T chip has doubled from the previous generation's 125 PFLOPS to 250 PFLOPS, and the memory bandwidth has doubled from 21.6 PB/s to 43.2 PB/s. Cerebras claims that the CS-4 has set a new industry benchmark in inference speed. In direct comparison tests on the GPT-OSS-120B model, the CS-4 can generate over 4,400 tokens per user per second, reaching up to 30 times faster than GPU solutions. The system can support models with over 50 trillion parameters. Compared to the previous generation CS-3, CS-4's speed has doubled, with a throughput improvement of 10 times per watt, significantly enhancing the economic efficiency of data centers.