VDI: GRID vPC tested on a server with 2x Intel Xeon Gold 6148 (20c, 2.4 GHz), GRID vPC with T4-1B (64 VMs), VMware ESXi 6.7, NVIDIA vGPU Software (410.91/412.16), Windows 10 (1803), 2 vCPUs, 4GB RAM, Resolution 1920x1080 resolution, Single Monitor, VMware Horizon 7.6 User experience was measured using an NVIDIA internal benchmarking tool which measured remoted frames running office productivity applications such as Microsoft PowerPoint, Word, Excel, Chrome, PDF viewing and video playback.
Machine Learning: CPU nodes (61 GB of memory, 8 vCPUs, 64-bit platform), Apache Spark. 200 GB CSV dataset; Data preparation includes joins, variable transformations. GPU Server Config: Dual-Socket Xeon E5-2698 v4@3.6GHz, 20 T4 GPUs on 5 nodes, each with 4 T4 GPUs. All run on InfiniBand network, CPU data for XGBoost and Data Conversion steps are estimated based on measured data for 20 CPU nodes, and reducing execution time by 60% to normalize for training on smaller data set on T4.
Deep Learning Training and Inference: GPU: Dual-Socket Xeon E5-2698 v4@3.6GHz. GPU Servers: 2xT4s for Training, 1xT4 for Inference, NGC 18.11-py3 Container with CUDA 10.0.130; NCCL 2.3.7, cuDNN 7.4.1.5; cuBLAS 10.0.130 | NVIDIA Driver: 384.145.