NVIDIA AI Enterprise with VMware vSphere: Combining NVIDIA GPU’s Superior Performance, NVIDIA AI Software, and Virtualization Benefits for AI Workflows (Presented by VMware, Inc.)

NVIDIA AI Enterprise combines the superior power of NVIDIA GPUs and NVIDIA AI software with the virtualization benefits of VMware vSphere for AI workflows. We'll demonstrate the near-bare-metal performance of NVAIE in vSphere using industry standard benchmarks of MLPerf and standard ML workloads. We'll highlight performance results for multi-GPU configuration connected by NVLINKs and address trade-offs for multi-instance GPUs (MIGs — spatial sharing) and virtual GPU (vGPU — time-sharing). We'll provide best practices for distributed training using RDMA and NVIDIA NICs. We'll show End2End AI workflow using Tanzu Kubernetes Grid (TKG) and best practices for training and tuning the ML model, deploying the model in production for inferencing using NVIDIA TensorRT, and scaling and production deployment. We'll discuss how to improve latency, throughput, and multi-tenancy by mixed-precision TensorCores, NVIDIA TensorRT, and vGPUs. Finally, we'll present our recent work on network function virtualization using vGPUs and MIGs.