NVIDIA’s DGX-1 Deep Learning appliance is a purpose-built deep learning and AI-accelerated platform, that delivers performance equal to approximately 250 conventional servers. It is based on the latest Pascal architecture-class of NVIDIA GPUs and four Mellanox ConnectX®-4 EDR 100Gb/s InfiniBand HCAs. DGX-1 leverages multiple EDR InfiniBand ports to eliminate performance bottlenecks and to provide maximum data throughput at ultra-low latencies. Leveraging the network acceleration capabilities for data offloads and collective communications, NVIDIA has demonstrated multi-system scalability of the DGX-1 platform by connecting 124 DGX-1 systems forming the SaturnV Machine Learning Supercomputer. SaturnV achieved not only the #1 spot on the Green500 list, but also captured the #28 spot on the Top500 with 3.3 petaflops. The multi-system DGX-1 cluster uses a network topology based on fat-tree design with InfiniBand and provides 20X network performance than that of a typical multi-system cluster, ensuring the highest communication performance among computational and deep learning applications.