Over 11% of Meta's GPUs sit idle in Prometheus supercluster

Meta has reported that over 11% of the graphics processing units (GPUs) in its Prometheus supercluster, located in New Albany, Ohio, are either idle or have failed. This supercluster is one of the company's largest computing deployments, designed to keep pace with AI training, a critical area for its technological development.

The reasoning behind having a significant number of idle GPUs is to ensure that the system can quickly recover from failures. In an extensive system like Prometheus, which is designed to scale to a gigawatt of capacity with approximately a million accelerators, failures are unavoidable. By having additional GPUs ready to step in, Meta aims to minimize the impact on its operations.

This approach is not unique to Meta; many companies operating large computing clusters face the challenge of reliability. As the number of components grows, so does the likelihood of one or more failing, which can disrupt critical processes like training large language models (LLMs) or high-performance computing (HPC) workloads.

The Prometheus supercluster is powered by Nvidia's Grace Blackwell rack systems and will soon include Vera Rubin systems. This highlights the ongoing collaboration between Meta and Nvidia in the realm of artificial intelligence, a partnership that is likely to continue evolving as both companies seek to optimize their computing capabilities.

The decision to maintain a high number of idle GPUs underscores the importance of reliability in AI operations. For users and developers, this means that Meta is committed to providing continuous and efficient service, which could translate into faster advancements in artificial intelligence and other technological applications.