Challenge the GPU throne with 30x token speed! France and France praise Cerebras (CBRS.US) CS-4 for opening the “wafer-level+decoupled reasoning” era

The Zhitong Finance App learned that after the artificial intelligence (AI) chip company Cerebras Systems (CBRS.US) held the Supernova Day event on Tuesday, BNP Paribas said it was deeply impressed by the partners and related announcements announced by the company.

BNP Paribas analyst Karl Ackerman wrote in a note to clients: “Like NVDA.US (NVDA.US)'s LPU, Cerebras's new partnership with OpenAI, Amazon (AMZN.US) AWS, and AMD (AMZ.US) highlights the growing market demand for improved decoding that is highly sensitive to delays in the inference process.”

The analyst further stated that in terms of fast reasoning, the newly announced SRAM-based CS-4 wafer-scale engine (Wafer-Scale Engine) provides “an amazing 43.2 pB/s on-chip memory bandwidth” and is equipped with a novel wafer I/O interface using RDMA. Cerebras is proud to say that the CS-4 system generates tokens 30 times faster than GPU systems in current production environments.

Additionally, analysts said that Cerebras's product roadmap to double performance every year and plans to increase throughput by 20 times by 2027 is impressive. The analyst said, “Cerebras is focusing on decoupled inference, by positioning its wafer-level engine as an ultra-high-speed decoding engine to enable fast token generation and work in collaboration with compute-intensive systems such as AWS Trainium or AMD Helios. Cerebras said its decoupled inference system is expected to be up to 10 times faster than GPUs and 5 times faster than configurations using only WSE.

Finally, analysts said that Arista Networks (ANET.US) has become Cerebras' “key” network partner, and it is expected that this partnership will help Arista expand new customer application scenarios over time.

According to reports, Cerebras officially launched the next-generation rack-level AI system Cerebras CS-4 at the Supernova Day event on Tuesday, claiming that it is “the fastest AI accelerator in the industry.” The system targets large model inference circuits and directly targets Nvidia's GPU server products. Cerebras also announced partnerships with companies such as OpenAI, AMD, Arista Networks, AWS, Figma, Cognition, and CrowdStrike at the event.

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

Cerebras claims that the CS-4 has set a new industry benchmark in the speed of reasoning. In direct comparison tests on the GPT-OSS-120B model, CS-4 can generate more than 4,400 tokens per second per user, up to 30 times that of the GPU solution. The system can support models with more than 50 trillion parameters. Compared to the previous generation CS-3, the CS-4 is twice as fast and has a 10-fold increase in throughput per watt, significantly improving the economic efficiency of the data center.