CAIOS Achieves 7+ GB/s per GPU on NVIDIA Blackwell Ultra | CoreWeave

This post details new benchmark results for CoreWeave AI Object Storage (CAIOS) on NVIDIA Blackwell Ultra GPU nodes, achieving an average throughput of 7+ GB/s per GPU. This represents a >3x improvement over previous H100 benchmarks. The gains are attributed to: 1) fewer GPUs per node (4 on Blackwell Ultra vs. 8 on H100), doubling per-GPU throughput; 2) the use of NVIDIA Quantum InfiniBand transport, which improved throughput by approximately 50% compared to Ethernet; and 3) LOTA pipeline optimizations, resulting in an additional ~17% performance improvement. The benchmark setup used the Warp S3 benchmarking tool on CoreWeave Kubernetes Service (CKS) with 16 Blackwell Ultra nodes, testing both Ethernet (TCP) and RDMA (InfiniBand) transports.