MLPerf Storage adds two inference tests

MLCommons’ MLPerf Storage v3.0 drew 19 organizations and adds KV-cache and vector-database tests alongside training and checkpointing. The official results also add S3 object storage beside POSIX; on-premises submissions reached a median 14 GB per second per watt for checkpoint writes, but the benchmark remains a procurement input, not a universal workload proxy. Teams scaling long-context agents should benchmark storage and cache paths—not only GPUs—alongside the archive’s analysis of Nvidia’s rising share of hyperscaler capex.