MinIO ships tiny AIStor for BlueField‑4
MinIO released AIStor targeted at NVIDIA BlueField‑4 — an embedded storage stack built for AI inference that runs on ARM and ships as a sub‑200MB binary, aimed at high‑performance, low-footprint deployments (x.com). That makes it a neat fit for edge inference setups where storage overhead and latency matter. (x.com).
MinIO’s engineering blog post (Mar 16, 2026) describes S3-over-RDMA paths that push object data directly into GPU memory at “wire speed,” claiming a zero-translation data path to accelerate inference and retrieval workloads. (min.io) AIStor with NVIDIA GPUDirect RDMA for S3-compatible storage is available now as a tech preview, while MinIO says general availability is expected in the second half of 2026 with early access for qualified enterprise customers. (min.io) A MinIO benchmark showed a 106 million‑vector index build completing roughly 12x faster when AIStor was paired with Milvus and NVIDIA cuVS under GPU-accelerated, RDMA-enabled configurations. (min.io) MinIO positions AIStor as a fit for the NVIDIA STX rack-scale reference architecture and says the software can saturate 800GbE network fabric to feed accelerators without sitting behind a host CPU. (storagenewsletter.com) The company describes AIStor as fully disaggregated with a distributed namespace designed to scale to multiple exabytes, removing centralized metadata tiers that can bottleneck growth. (storagenewsletter.com) MinIO maintains AIStor artifacts and docs on GitHub and highlights native support for Iceberg tables, Delta Sharing, and zero‑copy access patterns for platforms like Databricks as part of its integration roadmap. (github.com)
