MLPerf Storage
DPU-accelerated NVMe/TCP storage for AI training
Led MangoBoost’s MLPerf Storage benchmark submissions.
In v2.0 (2025), Mango StorageBoost sustained line-rate throughput over a 400G Ethernet fabric, delivering 6.2x GPU scalability and 1.57–2.05x higher throughput per unit of network bandwidth over competing solutions, and outperforming NVIDIA BlueField-3 in both NVMe/TCP and NVMe/RDMA configurations.
In v1.0 (2024), the submission averaged 2.7x (up to 4.7x) higher performance than other submitters on ResNet50 and 3D-UNet.