MinIO: you don't need an object cache if your object store is fast enough

MinIO compared CoreWeave's LOTA caching proxy benchmark with its own AIStor object storage. LOTA on 20 GPU nodes hit 18.4 GiB/s per node, while AIStor on Solidigm QLC SSDs with no cache reached 33.5 GiB/s per node. MinIO argues a fast erasure-coded object store can outperform a local NVMe cache.

• CoreWeave LOTA: 18.4 GiB/s per node with 1 TiB cache and dual 100 Gbps links
• MinIO AIStor without cache: 33.5 GiB/s per node on Solidigm QLC SSDs
• AIStor beat CoreWeave's cold path by more than 11x at cluster level
• MinIO develops its own MemKV store on raw NVMe for inference