MLPerf Storage v3.0 Adds Native S3 Support, Validating Object Storage for AI Workloads

MLCommons released the official benchmark results for MLPerf Storage v3.0, expanding its standardized suite to measure how modern storage architectures handle production machine learning workloads. The most significant architectural change in version 3.0 is the introduction of a native S3 object storage access layer alongside the traditional POSIX-compliant interface. In this benchmark cycle, approximately one-sixth of the total submitted configurations evaluated training and large-model checkpointing directly against S3-compatible object storage backends, alongside newly introduced test suites for LLM key-value (KV) caching and vector database indexing. For platform engineers and AI infrastructure architects, this benchmark marks a pivotal shift in validating object storage for high-performance computing. Historically, the prevailing assumption in distributed training was that only parallel file systems—such as Lustre, GPFS, or specialized NVMe-over-Fabrics appliances—could sustain the strict throughput and ultra-low latency required to keep modern GPU accelerators from starving. The lack of standard object storage benchmarks made it difficult to justify running direct-to-object pipelines. By establishing standardized metrics using the S3 API across multi-node configurations, infrastructure teams now have empirical, vendor-neutral data to gauge accelerator utilization and determine when object-native architectures are viable. This development aligns with the broader cloud-native movement toward disaggregating compute and storage while eliminating unnecessary data replication. In enterprise environments, unstructured data lakes and multimodal datasets natively reside in object stores like Amazon S3, Google Cloud Storage, or Azure Blob Storage. Forcing datasets through intermediate caching appliances or copying petabytes to transient scratch filesystems adds massive operational overhead, synchronization lag, and cost. As object storage engines improve horizontal parallelism and integrate fast NVMe tiers, modern cloud storage layers are increasingly capable of directly feeding multi-gigabyte-per-second training and checkpointing pipelines. In practice, engineering teams should leverage the MLPerf Storage v3.0 results to audit their storage procurement and architecture strategies. The benchmark demonstrates that high-throughput training and large-model checkpointing (such as Llama 3 checkpoint runs) can achieve required accelerator utilization thresholds directly over S3 endpoints. However, teams must evaluate workload characteristics carefully: while streaming large sequential training batches over S3 performs exceptionally well, random-access latency in inference-time KV caching and vector search still demands optimized caching layers. Organizations should evaluate these benchmarks to right-size their storage tiers, balancing the operational simplicity and cost benefits of S3-native infrastructure against raw POSIX filesystem throughput.