The GPUltima is a high-density, fully integrated computer cluster that is purpose-built for high performance computing (HPC) applications like financial trading, deep learning or machine learning, oil and gas exploration, virtual desktop infrastructure (VDI), defense and security, and academia. Where conventional computer cluster systems use CPUs as the primary data processor, the GPUltima employs numbers of GPU cards, providing 10 times the performance by adding thousands more cores. In addition, the GPUltima consumes about 90% less power than conventional systems and occupies 95% less rack space. The GPUltima petaflop edition requires only a single 42U rack and requires only 56kW of power. The GPUltima is completely 'application-ready', so that all the customer has to do is to add his application software to the servers and the system is ready to begin processing. The unique cluster management and monitoring software and the service and support packages that accompany the GPUltima make this a user-friendly system that allows the customer to begin his work without having to configure the cluster. The GPUltima is comprised of 'compute nodes'. Each compute node contains sixteen GPU cards and one or two dual socket servers, each with dual 'Haswell' processors, producing 139 Teraflops of performance using NVIDIA Tesla K80s. The sixteen GPU cards communicate with each other through Infiniband EDR 100Gbs connections through a 1U Infiniband switch. The server(s) are cabled to the GPU enclosure through 128Gbs PCIe. The server communicates to the internet through Ethernet. Additional nodes can be added as needed. For more than one server an Ethernet switch is added for external communication. The clustering software allows complete manageability of the nodes as well as the individual GPU cards. The GPU monitoring and management software provides a 'single-pane-of-glass' management of the hardware, the operating system, HPC software, and users. With the Cluster Manager, system administrators can quickly get clusters up and running and keep them running reliably throughout their lifecycle - all with the ease and elegance of a full-featured, enterprise-grade cluster manager.
