AI hardware innovator Cerebras Systems has officially unveiled Condor Galaxy 3, a colossal AI supercomputer designed to challenge the industry’s reliance on GPU-based infrastructure. The new system, located in Dallas, Texas, delivers a staggering 64 exaflops of AI compute power, positioning it as one of the most powerful AI training platforms in the world and a direct competitor to Nvidia’s dominant hardware ecosystem.
Constructed from 64 interconnected Cerebras CS-3 systems, each powered by the company’s signature Wafer-Scale Engine 3 (WSE-3) chip, Condor Galaxy 3 is built for simplicity and scale. Unlike traditional supercomputers that require complex networking of thousands of individual GPUs, the Cerebras architecture is designed to make training massive, multi-trillion-parameter models as straightforward as training a model on a single device. This approach aims to eliminate the significant software and engineering challenges associated with distributed computing across vast GPU clusters.
Developed in partnership with Abu Dhabi-based technology group G42, this supercomputer is the third installment in their Condor Galaxy network. Cerebras’s technology allows for what it calls “weight streaming,” enabling models with trillions of parameters to be stored off-chip in system memory while being efficiently streamed to the wafer-scale processor for computation. This eliminates bottlenecks and simplifies the development process for organizations building frontier AI models.
The launch of Condor Galaxy 3 marks a significant move in the diversification of the AI hardware market. As demand for AI compute continues to explode, companies are actively seeking viable alternatives to Nvidia’s often supply-constrained GPUs. By offering a purpose-built, highly scalable, and user-friendly system, Cerebras is making a compelling case for a new architectural approach to powering the next generation of artificial intelligence.
English 
Deutsch