Storage for AI is rapidly changing: Checkpointing becomes more important as clusters scale to more accelerators, managing large KV-Caches from LLM queries shifts inference bottlenecks to storage, accessing relevant data via VectorDB similarity searches drives small IOs for nearly every query, and future applications may require wildly different storage architectures

The MLPerf Storage v2.0 Benchmark Results were just released and the v2.5 suite is under active development. In this session we go over the Checkpointing benchmark added to v2.0 and the KV Cache Management and VectorDB benchmarks planned for v2.5. We will analyze IO traces of the workloads to understand how these workloads translate from the application layer to storage and discuss how these workloads may drive future storage performance requirements.