Case Study: Meta accelerates Llama 3 training and large-scale GenAI development with Hammerspace

Meta - Customer Case Study

Meta needed storage that could support its rapidly expanding GenAI training environment, including Llama 3, across two 24,576-GPU clusters. The company faced the challenge of enabling fast checkpointing, high-throughput data loading, and an improved developer experience at extreme scale, and partnered with Hammerspace to add a parallel network file system for its AI workloads.

Hammerspace worked with Meta to co-develop and deploy an NFS solution alongside Meta’s Tectonic distributed storage, helping engineers access code changes immediately across thousands of GPUs for faster interactive debugging and iteration. The combined storage approach improved usability without sacrificing scale, supporting synchronized checkpointing and exabyte-scale data access while helping Meta sustain high-performance training on its large RoCE and InfiniBand clusters.

Meta