Paper Title: 1PB of NVMe-oF Flash Storage for AI

Paper Abstract: As featured on Linus Tech Tips, NVIDIA, KIOXIA, AMD, Supermicro and Weka collaborated to bring together a high-performance PB of PCIe 4.0 NVMe flash storage for AI applications. this configuration consisted of NVIDIA A100 processors in a Supermicro HGX server as a load generation platform, connected via eight 200GbE network ports to six storage servers each with 184.32TB of PCIe 4.0 NVMe SSDs and two 200GbE network ports each. NVMe-oF over RDMA protocol was used to transfer data over the network using GPUDirect Storage technology to power the AI applications. This session looks at the real-world creation, challenges and successful deployment of the configuration, and the application benefits to AI.

Paper Author: Rob Davis, VP Storage Technology - Networking Business Unit, NVIDIA

Author Bio: Rob Davis is the VP of Storage Technology at the NVIDIA Networking Platform Group enabling ways to apply high-speed interface technology to storage systems. For more than 35 years, he has been a key contributor and decision maker for the development of storage networking products, including products based on NVMe over Fabrics technology. Rob was previously VP and CTO Officer at QLogic, where for 15 years he drove development and marketing of Fibre Channel, Ethernet and InfiniBand technologies. His in-depth expertise spans virtualization, Ethernet, Fibre Channel, SCSI, iSCSI, InfiniBand, RoCE, SAS, PCI, PM, SATA, and flash Storage