“Novo-G: A View at the HPC Crossroads for Scientific Computing” ERSA Keynote for Reconfigurable Supercomputing Panel A. George, H. Lam, C. Pascoe, A. Lawande, G. Stitt NSF Center for High-Performance Reconfigurable Computing (CHREC) ECE Department, University of Florida, Gainesville, FL 32611-6200 {george, hlam, pascoe, lawande, gstitt}@chrec.org Abstract - High-performance computing for many science domains is at a major crossroads. Conventional HPC is ill- equipped to address escalating performance demands without resorting to massively large, energy-hungry, and expensive machines. This paper focuses upon the principal challenges for HPC for scientific computing, why and how reconfigurable supercomputing is poised to make a major impact on accelerating scientific applications. Novo-G, recently fielded by CHREC, the NSF Center for High-Performance Reconfigurable Computing, is believed to be the most powerful reconfigurable computer in the research community. Novo-G is an experimental testbed that serves as the centerpiece for a variety of research projects dedicated to understanding and advancing performance, productivity, and sustainability of future systems and applications. It is also an integration point for the Novo-G Forum, comprised of an international group of academic researchers and technology providers working collaboratively on applications and tools to establish and showcase advantages of reconfigurable computing at scale via Novo-G. Results from impactful applications and highlights from research projects being adapted for Novo-G by CHREC at the University of Florida will be presented. Keywords - reconfigurable supercomputing, reconfigurable computing (RC), high-performance computing (HPC), FPGA, scientific computing 1. Introduction High-performance computing for many science domains is at a major crossroads and in the center of a convergence of several technology megatrends of the last decade. First, technological advances in these areas have transformed data- starved science domains into data-driven ones. One such high-profile example is the Large Hadron Collider (LHC) project from CERN (European Organization for Nuclear Research). According to CERN’s Web site [1], the LHC will produce roughly 15 Petabytes (15 PB) of data annually. Another example is the Large Synoptic Survey Telescope (LSST) project, which is scheduled to see first light in 2014, to begin doing science in 2015, and be in full survey operations by 2016. It is projected to produce 30 Terabytes (TB) of data per night, leading to a total database over the ten years of operations of 60 PB for the raw data, and 30 PB for the catalog database [2]. In the domain of genomics research, today’s DNA sequencing instruments are capable of producing 225-250 million DNA reads per run, resulting in output files routinely in excess of 1 TB per instrument run. In the near future, instruments are projected to produce up to 3 billion reads per run, in which the output of a single run will approach 100-fold coverage of the entire human genome. Thus, it is with increasing recognition that the discordant trajectories growing between data production and capacity for timely analysis is threatening to impede new scientific discoveries and progress in many scientific domains, not because we cannot generate the raw data, but because we cannot analyze it. In the second technology megatrend during the last decade, conventional computing no longer can depend upon exploiting increased clock rate and instruction-level parallelism to sustain performance. Performance is now improved through explicit parallelism using multicore and manycore processors. As a result, conventional HPC is ill- equipped to address escalating performance demands without resorting to massively large, energy-hungry, and expensive machines, where designers simply throw thousands (and soon millions) of x86 processor cores at each new and demanding problem. Lastly, reconfigurable computing (RC) is finally ready for prime time. The research community has demonstrated small-scale but exciting successes in applying FPGA-based RC technology for accelerating scientific applications. Today’s (and emerging) FPGAs finally have the computational horsepower (upper hundreds of thousands logic elements, soon millions, with tens of millions of memory bits, and other built-in functions) to enable high- performance computing. The opportunity is ripe for reconfigurable supercomputing (a.k.a., RC supercomputing), using scalable RC systems featuring a relatively large number of leading-edge FPGAs that can be configured specifically for high-intensity data processing for each application. The NSF Center for High-Performance Reconfigurable Computing (CHREC) recently fielded what is believed to be the most powerful reconfigurable computer in the research community. This machine, called Novo-G, is an experimental testbed that serves as centerpiece for a variety of research projects dedicated to understanding and advancing performance, productivity, and sustainability of future systems and applications. Novo-G features 192 new and extremely powerful FPGA accelerators, as well as potent subsystem and system architectures. The remainder of the paper is organized as follows. A discussion of the three principal challenges for HPC for scientific computing, performance, sustainability, and productivity, is given in Section II. In Section III, the architecture and features of Novo-G are presented, along with the goals and organization of a Novo-G Forum. Sections IV and V highlight Novo-G research activities that