The Weekly Top Five
The Weekly Top Five features the five biggest HPC stories of the week, condensed for your reading pleasure. This week, we cover the Cray-Sandia partership to found a knowledge institute; RenderStream’s FireStream-based workstations and servers; NVIDIA’s CUDA center growth; Reservoir Labs and Intel’s extreme scale ambitions; and Jülich Supercomputing Centre’s new hybrid cluster. Plus a bonus section.
Cray, Sandia Combine Efforts to Foster Knowledge Discovery
Prominent supercomputer vendor Cray Inc. and Sandia National Laboratories have come together to establish the Supercomputing Institute for Learning and Knowledge Systems (SILKS). This is a Cooperative Research and Development Agreement (CRADA), a private-public collaboration, which aims to promote knowledge discovery, data management and informatics computing. SILKS is located at Sandia’s Albuquerque-based headquarters and draws from its founders’ hardware and software resources as well as the experience and knowledge of their research staff.
The founding partners declared three primary goals for the endeavor:
1. Accelerate the development and application of high performance computing (HPC) technologies focused on solving knowledge discovery, data management and informatics problems at scale.
2. Collaborate to overcome the implementation barriers to a wider adoption of data-driven HPC computing technologies in knowledge discovery, data management and informatics.
3. Apply the use of these technologies to enable discovery and innovation in science, engineering and for homeland security.
The broad-based agenda will tackle a range of technology domains, including software, hardware, services, education and outreach. Representatives from both Sandia and Cray anticipate the collaborative effort will result in cutting-edge technologies and solutions.
RenderStream Releases AMD-based Servers and Workstations for OpenCL
Addressing the need for GPU-accelerated HPC, 3D workstation specialist RenderStream has launched its AMD Radeon-based servers and workstations for OpenCL. These 21.6 teraflop systems also support OpenGL and Brooks based applications and product development. According to a company statement, GPGPU high-performance computing using AMD GPUs shows great potential for information security, medical imaging, computer graphics and rendering, server side rendering, finite-difference-time-domain (FDTD), electro-magnetics, physics, bio-science and EDA.”
RenderStream’s AMD Radeon HD 6970 based VDACTr8 and its HD 6990 based VDAC4x2 implement 1,536 stream processors and eight GPUs per system, providing access to 12,288 cores and 21.6 teraflops of computing power when operating at an over-clocked peak performance.
The official announcement illustrates the server’s performance-boosting capabilities with this real-world example from the field of information security:
Using the integer-based oclHashCat, RenderStream’s customers are seeing near linear scaling in computational power which simply trounces the 4,096 cores and 12.6 teraflops of our GTX 580 based VDACTr8. In this example the HD 6970 and HD 6990 based VDACTr8 evaluated over 45 billion solutions per second versus 18 billion for the GTX 580 based systems, depending on the implementation.
RenderStream offers general purpose GPU systems as well as HPC-specific GPU-based platforms outfitted with either NVIDIA Tesla or AMD FireStream graphics processors.
NVIDIA CUDA Centers Number Four Hundred
This week, NVIDIA announced the addition of 35 new CUDA Research Centers and CUDA Teaching Centers, bringing the total number of such centers to 400. The latest partner institutions come from 14 countries, evidence of parallel computing’s — and NVIDIA’s — global reach.
The centers will leverage the parallel computing power of NVIDIA’s CUDA-based GPUs to tackle a bevy of challenging computing issues, as well as teach thousands of students cutting-edge GPU programming skills. CUDA Research Centers employ GPU computing across multiple domains, while the CUDA Teaching Centers have incorporated GPU computing techniques into their main computer programming curriculum. NVIDIA explains that its CUDA Research Center Program “fosters collaboration at institutions that are expanding the frontier of parallel computing.” Partners benefit from “exclusive events with key researchers and academics, a designated NVIDIA technical liaison, and access to specialized online and in-person training sessions.”
For a full listing of the newest CUDA Research Centers and CUDA Teaching Centers, see the official announcement.
Reservoir Labs, Intel Partner on DARPA UHPC Project
Reservoir Labs announced it will collaborate with Intel researchers on the development of compiler technologies and architectures in order to create viable extreme scale computing by the year 2018. The duo have signed a subcontracting agreement that brings Reservoir Labs research scientists and technologies to Intel’s team to develop Extreme Scale computing technologies as part of DARPA’s Ubiquitous High Performance Computing (UHPC) research program.
According to the release: “The goal of the UHPC program is to develop 1 PFLOPS (HPL) single cabinet systems, including self-contained cooling, that overcome significant energy efficiency, security, and programmability challenges. Essentially this can be viewed as integrating the computational capacity of today’s largest supercomputers in 100x less area, with 100x less power, and with significant increases in programmability and applicability.”
Intel’s UHPC team is tasked with supporting and developing technologies to enable the US to build extreme scale computers by the year 2018. In order for this challenging goal to come to fruition, major breakthroughs in hardware and software design will be necessary, far beyond the level of current commercial offerings. Just improving the energy efficiency levels of computers by more than 100x will require significant advancements.
If these goals are achieved, the resulting technology would benefit embedded applications, such as those found in ship, land, and air-based Department of Defense systems. Exteme scale systems would also further other DoD objectives, such as Intelligence Surveillance Reconnaissance (ISR), Electronic Warfare (EW), Integrated Air and Missile Defense (IAMD), battle management and planning, and cyber security.
The initial contract calls for the project to furnish a “proof of concept” implementing extreme scale technologies in a first-pass system design by 2012. A second phase is also outlined, which if DARPA elects to continue, could lead to a completed system design for 2014 timeframe. The full scope of the contract specifies the delivery of a prototype extreme scale system in 2018.
Jülich Supercomputing Centre Debuts Hybrid System
A new GPU-accelerated system will support high-level research at the Jülich Supercomputing Centre (JSC) in Germany. The hybrid cluster, named JUDGE, for “Jülich Dedicated GPU Environment,” relies on GPUs to boost processing power, while minimizing energy consumption. JUDGE will be used for data-intensive workloads in the fields of biology, medicine and environmental research.
The cluster was built using 54 IBM System x iDataPlex server nodes with 12 cores each and 96 GB memory, as well as 108 NVIDIA M2050 GPUs. The release describes IBM iDataPlex as “a scalable system that can significantly reduce energy consumption, cooling and space requirements.”
Martin Hiegl, the team leader for Deep Computing Sales at IBM Germany, commented, “Together with JSC’s other powerful supercomputers, the new JUDGE cluster supports Germany’s ability to tackle a wide range of scientific and technical challenges.”
Sales leader for HPC at NVIDIA, Stefan Kraemer, believes the hybrid design, which relies on the GPU’s accelerative force, will be the template for the coming exascale generation. “The JUDGE cluster is a good example of how we need to continue to develop computers in the future, following the target of exascale computing. This is valid not only in regard to performance, but also to energy consumption and energy efficiency,” he states, adding: “Pilot projects like JUDGE play a key role in this process and are a key step on the way to hybrid systems.”
JUDGE is not the first IBM/JSC collaboration. The duo united to create the QPACE supercomputer, which consistently ranks among the top ten of the Green500 list of the world’s most energy efficient supercomputers, and also worked together on the Blue Gene/P-based JUGENE, one the most powerful computers in Europe with a peak performance of more than one petaflop.
There was such a grand allotment of noteworthy news this week that we are presenting our first ever bonus link section: