SEATTLE— The Barcelona Supercomputing Center (BSC) has announced it will develop a hybrid supercomputer based on Nvidia Corp.'s Tegra ARM CPUs and the firm's CUDA-supporting Tesla GPUs, with hopes of reaching exascale performance.
BSC believes its prototype system will be the world's first ARM-based CPU/GPU supercomputing combination. The center says it is aiming for a two to five times improvement in energy efficiency compared to today's most efficient systems in the short term, with the ultimate goal of reaching exascale using 15 to 30 times less power than current supercomputer architectures.
The lofty goal has been aptly dubbed the EU Mont-Blanc Project and is mainly being carried out as a proof of concept, to show what might be possible in future on more energy-efficient, embedded mobile technologies and jumpstart software development for the ARM architecture in the space.
"In most current systems, CPUs alone consume the lion's share of the energy, often 40 percent or more," said Alex Ramirez, leader of the Mont-Blanc Project. "By comparison, the Mont-Blanc architecture will rely on energy-efficient compute accelerators and ARM processors used in embedded and mobile devices to achieve a four- to 10-times increase in energy-efficiency by 2014."
Nvidia's director of Tesla marketing, Sumit Gupta, said the idea of ARM playing in the supercomputing space was not as far-fetched as some would believe, however, especially not now that ARM CPUs are already being tried out experimentally in cloud servers. Gupta alluded to the importance of Calxeda's initiative with Hewlett Packard Co., important not just in terms of concept, but because HP's sales volumes are typically large.
"ARM is going to happen, no matter what people want or like," said Gupta. "ARM is the future for HPC and the PC."
While using graphics processors is still a fairly young technology in the supercomputing space, said Gupta, it is starting to see high levels of adoption already, and the software ecosystem is starting to emerge around it.
To speed up the trend and encourage more ARM-based initiatives, Nvidia has also said it plans to develop a hardware and software development kit which will feature a quad-core Nvidia Tegra 3 ARM CPU accelerated by a discrete Nvidia GPU.
The kit's hardware will be developed by SECO and is expected to be available in the first half of 2012. Nvidia says it will support the hardware with its own proprietary CUDA parallel programming toolkit.
So grateful is Nvidia to HPC centers using GPU technology that the firm has even devised an incentivized reward scheme, handing out the title of "CUDA Center of Excellence."
BSC has recently been granted the title, as has Lomonosov Moscow State University, joining 13 others including John Hopkins University, Stanford University, Harvard University, Institute of Process Engineering at the Chinese Academy of Sciences, National Taiwan University, Tokyo Tech, Tsinghua University (China), University of Cambridge, University of Illinois at Urbana-Champaign, University of Maryland, University of Tennessee, Georgia Tech, and University of Utah.
BSC, a high-performance computing research center associated with the Universitat Politecnica de Catalunya/Barcelona Tech, is Spain’s national supercomputing facility and home to one of Europe’s most powerful supercomputers, the MareNostrum. The institute also recently deployed Spain's fastest compute cluster with 256 Nvidia Tesla M2090 GPUs and quad-core CPUs, said to deliver a peak performance of 186 teraflops.
Why GPU? Why not something that specially made for parallel processing? GPU seems a little odd in terms of computing but interestingly some people changes its use to become a faster processing tool. Want to know more about the technology behind.
I am intrigued by the thought of GPU based supercomputers, but wonder if the compilers are up to the task? Is this the start of something that is a redo of older technology? I am thinking of the original number crunchers like the Cray and other array processors that were hardware specialized machines with fast execution times built on pipelined architectures. Seems to me that GPUs being specialized cores for graphics with pipelined for performance constructs are not that different from the "old technology" of array processors. I wonder why it took so long, maybe it is related to my original question about compiler technology?
Join our online Radio Show on Friday 11th July starting at 2:00pm Eastern, when EETimes editor of all things fun and interesting, Max Maxfield, and embedded systems expert, Jack Ganssle, will debate as to just what is, and is not, and embedded system.