Cerebras Systems has launched its latest AI accelerator, the CS-4, which the company says can deliver inference speeds up to 30 times faster than existing GPU-based solutions.
The CS-4 is described as the first implementation of the company’s new Nexus rack-scale platform and features several new hardware and architectural developments aimed at supporting large-scale AI models.
Go deeper with GlobalData
Access deeper industry intelligence
Experience unmatched clarity with a single platform that combines unique data, AI, and human expertise.
The newly introduced CS-4 integrates three Wafer Scale Engine 3 Turbo (WSE-3T) processors, each containing four trillion transistors and 900,000 AI-optimised cores over a silicon area of 46,225mm².
With a total of 750 PFLOPs of AI compute, the CS-4 supports a memory bandwidth of 129.6 petabytes per second across the system, facilitated by a bandwidth of 7.2 terabits per second for input and output.
Cerebras states that the CS-4 produces up to twice the performance of its predecessor, the CS-3, with improvements including up to ten times more throughput per watt.
According to the company, these efficiency gains are designed to improve data centre operations by allowing both higher processing speeds and increased token generation within the same power constraints.
Cerebras CEO and co-founder Andrew Feldman said: “In AI, speed is productivity. Historically, fast inference meant using smaller and less capable models. Cerebras CS-4 delivers industry-leading speeds on the largest frontier models, fundamentally changing the paradigm.
“Every aspect of the design has been optimised to deliver the highest speeds with massive throughput. With the CS-4, AI is so fast that it fundamentally reshapes product experiences.”
The CS-4’s modular rack design, built upon the Nexus Platform Architecture, incorporates three primary elements: compute capability, power delivery, and input/output (I/O).
Innovations highlighted by Cerebras include a rear-mounted “Wafer-Scale Backpack” for integrating power, cooling, I/O, and control elements around the wafer, which the firm says halves the number of components and streamlines deployment from days to hours.
Supporting models with over 50 trillion parameters is possible due to reduced wafer-to-wafer latency, claimed to be as low as two microseconds, said the US-based AI hardware company.
New features also include a programmable I/O subsystem, supporting both standards-based RoCE v2 RDMA over Ethernet and a direct wafer linking technology, and what the company describes as a high-density power delivery system that reduces energy losses.
Cerebras chief technology officer (CTO) and co-founder Sean Lie said: “Being 30 times faster doesn’t just make a response feel fast. It gives an agentic system room for more than an order of magnitude as much reasoning, verification, or tool use in the same wall-clock time. That’s the difference CS-4 makes for real production workloads.”
Initial shipments of the CS-4 are scheduled to begin this quarter.
In July, Cerebras formed a partnership with CrowdStrike aimed at enhancing enterprise threat detection and response by integrating cybersecurity tools with rapid AI inference capabilities.
