30x Token Speed Challenges GPU Dominance! BNP Paribas Praises Cerebras (CBRS.US) CS-4 for Ushering in the "Wafer-Scale + Decoupled Inference" Era

According to Golden Ten Data, after artificial intelligence (AI) chip company Cerebras Systems (CBRS.US) held its Supernova Day event on Tuesday, BNP Paribas stated that the partners and related announcements revealed by the company left a strong impression.

BNP Paribas analyst Karl Ackerman wrote in a report to clients: "Similar to NVIDIA’s (NVDA.US) LPU, Cerebras’ new partnerships with OpenAI, Amazon (AMZN.US) AWS, and AMD (AMZ.US) highlight the growing market demand for improving the highly latency-sensitive decoding phase of inference processes."

The analyst further noted that for rapid inference, the newly unveiled SRAM-based CS-4 Wafer-Scale Engine offers an "astounding 43.2 PB/s on-chip memory bandwidth" paired with a novel wafer I/O interface leveraging RDMA. Cerebras proudly claimed that the CS-4 system generates tokens 30 times faster than GPU systems in current production environments.

Additionally, the analyst remarked that Cerebras’ roadmap of doubling performance annually with plans to increase throughput by 20 times by 2027 is admirable. He stated: "Cerebras is focused on decoupled inference, positioning its Wafer-Scale Engine as an ultra-fast decoding engine for rapid token generation, working in tandem with computation-intensive systems like AWS Trainium or AMD Helios. Cerebras says its decoupled inference system is expected to be up to 10 times faster than GPUs, and delivers 5 times 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 applications.

It is reported that Cerebras officially launched its next-generation rack-level AI system, Cerebras CS-4, at the Supernova Day event on Tuesday, calling it "the industry's fastest AI accelerator". The system is aimed at large model inference and directly targets NVIDIA’s GPU server offerings. At the event, Cerebras also announced partnerships with OpenAI, AMD, Arista Networks, AWS, Figma, Cognition, and CrowdStrike.

CS-4 is the first product in Cerebras’ new-generation Nexus rack-level platform architecture, powered by three brand-new Wafer Scale Engine 3 Turbo (WSE-3T) wafer-level processors. WSE-3T is currently the largest AI processor ever built, with each chip integrating 4 trillion transistors and 900,000 AI-optimized cores, spanning a silicon area of 46,225 square millimeters, with 44GB of built-in SRAM. The CS-4 system provides 750 PFLOPs of AI computing power, memory bandwidth of 129.6 PB per second, and I/O throughput of 7.2 Tbps per second. AI computing power per single WSE-3T chip has doubled from the previous generation’s 125 PFLOPS to 250 PFLOPS, and memory bandwidth has doubled from 21.6 PB/s to 43.2 PB/s.

Cerebras claims that CS-4 sets a new industry benchmark for inference speed. In direct comparison tests on the GPT-OSS-120B model, CS-4 can generate over 4,400 tokens per user per second, achieving up to 30 times the performance of GPU solutions. The system supports models with over 50 trillion parameters. Compared to the previous CS-3, CS-4 is twice as fast, has 10 times the throughput per watt, and significantly improves the economics of data centers.