HokieSpeed debuted on the November 2011 Green500 List as the highest-ranked commodity supercomputer in the U.S. It is a GPU-accelerated supercomputer that achieved 120.4 TFLOPS in a thermal envelope of 126 kW. The total budget for the entire HokieSpeed supercomputer, including the in-situ visualization system is $1.4M.
Specs: 120.4 TFLOPS, 928.96 MFLOPS/watt, 209 compute nodes, 2508 Xeon cores, 418 NVIDIA 2050 Accelerators, 187264 GPU cores, 6270 GB System Memory, 24 + 6 GB Memory per node, ConnectX-2 QDR Infiniband interconnects.
Since the 1960s, the general-purpose processor (also known as the central processing unit or CPU) has served as the brains in computing instruments. For example, each of the 1,100 compute nodes in Virginia Tech’s System X supercomputer has a pair of homogeneous brains, i.e., two 2.3-GHz PowerPC 970FX CPUs. Recent trends, however, have exposed the CPU as a “jack of all (computing) trades, master of none,” thus giving rise to heterogeneous computing instruments with multiple types of brains, e.g., CPUs and graphics processing units (GPUs). Building on our team’s expertise in this area, we propose to acquire a versatile heterogeneous supercomputing instrument, in which each compute node consists of CPUs and GPUs. This transformative instrument will empower faculty, students, and staff across disciplines to tackle problems previously viewed as intractable or that required heroic efforts and significant domain-specific expertise to solve. For example, in 2007, though conventional wisdom believed that finding missing genes in 699 microbial genomes was computationally infeasible, PI Feng led a team of 20+ interdiscipinary researchers from 8 institutions around the world and developed software cybertool instruments to integrate a set of distributed supercomputers, totaling 12,000+ CPUs, to complete the task in ten weeks. We anticipate that the proposed instrument, coupled with our existing cybertool instruments and those available from NVIDIA, will complete this task in a day while involving only 2 researchers rather than 20+.
Furthermore, this instrument will catalyze new approaches for conducting research via the synergistic amalgamation of heterogeneous supercomputing and our cyber-enabled tools that enhance ease of use. In particular, it will allow end users to (1) make commonplace, the ability to perform in-situ visualization for rapid visual information synthesis and analysis and (2) to control their level of immersion in the discovery process — from being completely immersed, a la a “human in the loop” making real-time decisions (intuitively) via a large-scale gigapixel display, to observing the instrument automatically collect, organize, and analyze data in support of visual analytics.