AI benchmarking consortium MLCommons expanded MLPerf Storage with key-value cache, vector-database, S3 object-storage, and POSIX file-system workloads.

MLCommons expanded the MLPerf Storage benchmark to test key-value cache and vector-database workloads. These workloads represent storage demands associated with large-scale AI inference and retrieval systems.

The suite also added support for Amazon S3-compatible object storage alongside its existing support for Portable Operating System Interface, or POSIX, file systems. This broadens the benchmark across object-storage and conventional file-system architectures.

MLCommons also released MLPerf Storage v3.0 results. The organization describes the suite as architecture-neutral, representative and reproducible, and said version 3.0 expanded coverage across the full range of AI storage workloads.

The benchmark measures storage-system performance for machine-learning workloads rather than GPU arithmetic performance or model quality. Its new tests therefore focus on how storage platforms handle data-access patterns associated with inference caches and vector databases.