Cerebras Systems has unveiled its Wafer Scale Engine 3 (WSE-3), dubbed “the world’s fastest AI chip.”
The WSE-3, which powers the Cerebras CS-3 AI supercomputer, offers twice the performance of its predecessor, the WSE-2, with the same power consumption and price.
The chip is capable of training AI models with up to 24 trillion parameters, a significant jump from previous models.
cs-3 supercomputer
The WSE-3 is based on a 5nm TSMC process and features 44GB of on-chip SRAM. It features four trillion transistors and 900,000 AI-optimized computing cores, delivering a maximum AI performance of 125 petaflops, which is the theoretical equivalent of about 62 Nvidia H100 GPUs.
The CS-3 supercomputer, powered by WSE-3, is designed to train next-generation AI models that are 10 times larger than GPT-4 and Gemini. With up to 1.2 petabyte system memory, it can store 24 trillion parameter models in a single logical memory space, simplifying training workflow and increasing developer productivity.
Cerebras says its CS-3 supercomputer is optimized for enterprise and hyperscale needs, and offers superior power efficiency and software simplicity, requiring 97% less code than GPUs for large language models (LLM).
Cerebras CEO and co-founder Andrew Feldman said: “WSE-3 is the world's fastest AI chip, designed specifically for the latest cutting-edge AI work, from a combination of experts to 24 trillion-parameter models. “We are thrilled to bring WSE-3 and CS-3 to market to help solve today’s biggest AI challenges.”
The company says it already has a backlog of orders for the CS-3 across enterprise, government and international clouds. The CS-3 will also play an important role in the strategic partnership between Cerebras and G42, which has already delivered 8 exaFLOPs of AI supercomputer performance through Condor Galaxy 1 and 2. A third facility, Condor Galaxy 3, is currently under construction , which will be built with 64 CS-3 systems, producing 8 exaFLOPs of AI computing.
