Inside Prometheus: Meta's Journey to a Gigawatt-Scale Supercluster (engineering.fb.com)

Meta has unveiled Prometheus, its groundbreaking gigawatt-scale supercluster currently under construction in New Albany, Ohio, which aims to revolutionize the training of foundational AI models. The first phase is complete and already hosts hundreds of thousands of GPUs, with plans to scale up to approximately 1 million GPUs. This infrastructure is poised to enable the development of advanced AI capabilities that could enrich user experiences globally. Prometheus features innovative technologies such as closed-loop liquid cooling and a specialized networking architecture designed to maximize operational efficiency and minimize downtime. The significance of Prometheus lies in its potential to elevate AI model performance by providing unparalleled compute power and reliability. Notably, the supercluster uses advanced NVIDIA GPU platforms—Catalina and Clemente—which support high memory bandwidth, immense computational capabilities, and efficient interconnectivity through NVLink and Ethernet networks. These technical advancements allow for more extensive model training and increased resilience against computational failures, with provisions for rapid recovery from hardware issues. Ultimately, Prometheus represents a new era in AI infrastructure, capable of supporting the scale and complexity necessary for next-generation AI applications.