Nvidia is rolling out its new Tesla K20 family of GPU accelerators for high-performance computing (HPC) and supercomputing environments, offering the scientific community what officials said are more powerful and energy-efficient accelerators for their workloads.
Nvidia officials, who unveiled the GPU accelerators Nov. 12 at the SC12 supercomputing show in Salt Lake City, noted that the new K20X accelerators are key components of the Titan supercomputer at the Oak Ridge National Laboratory in Tennessee, which is now the world's fastest supercomputer with a performance of 17.59 petaflops (quadrillions of calculations per second).
Titan—an XK7 supercomputer from Cray—runs on Opteron 6274 processors from Advanced Micro Devices and Nvidia's GPU accelerators. In all, Titan holds 560,640 processors, including 261,632 of Nvidia's Tesla K20x GPU accelerators. Nvidia officials said 90 percent of Titan's performance comes from the GPU accelerators.
Accelerators increasingly are being used in supercomputers and HPC environments as researchers and organizations look for ways to boost the compute performance of their systems while driving down the amount of power they use. Nvidia and AMD over the past few years both have been aggressive in promoting GPU accelerators, which work with traditional CPUs to crank up system performance, particularly when crunching compute-intensive, highly parallel workloads. To that end, also at the supercomputer show, AMD unveiled its new FirePro s10000 GPU accelerators.
Intel is looking to offer systems makers and organizations an alternative with its upcoming x86-based Xeon Phi coprocessors, which company officials argue offer significant benefits over GPUs.
According to Sumit Gupta, general manager of Nvidia's Tesla Accelerated Computing unit, for a long time scientists felt that computing wasn't giving them the results they needed—in areas such as simulation, for example—to offset the more traditional experiments they were conducting. In addition, supercomputers traditionally were "power hogs," Gupta told eWEEK.
However, with the rise of GPU accelerators—Nvidia launched its Fermi accelerators in 2009—performance in supercomputers became better and power consumption was reduced, making supercomputers more powerful and less expensive to use.
"One of the things that has changed over the years is that the use of computing has become popular in scientific computing," he said. "We've seen the takeoff of the use of GPU accelerators.
According to the officials behind the Top500 list of the world's fastest supercomputers, 62 of the systems on the list now leverage GPU accelerators or coprocessors.
Since 2008, Nvidia has been on a two-year cycle of introducing new GPU architectures, with the new Kepler architecture—on which the K20 family is built—being the latest. In 2014, Nvidia officials expect to release the next generation, dubbed "Maxwell."
According to Nvidia officials, the new Tesla K20 family includes the K20 and K20X accelerators, the latter being the flagship chip in the family. The new GPUs offer three times the processing speeds of the previous-generation GPUs based on the Fermi architecture. The K2p offers 1.17 teraflops of peak double-precision performance, while the K20X provides 1.31 teraflops. At the same time, the new GPUs offer three times the energy efficiency of the prior accelerators, a key capability for HPC and supercomputing environments. According to Nvidia, with the K20X accelerators, the Titan supercomputer his 2,142.77 megaflops of performance per watt, which would rank it at the top of the most recent Green500 list of the world's most energy-efficient supercomputers.
Along with the introduction of the new chips, Nvidia's Gupta also stressed the combination of the accelerators with the CUDA programming model. CUDA is becoming increasingly popular, with 395 million CUDA GPUs having shipped and 1.5 million CUDA downloads. He also said that the programming language is being taught in 62 countries.
Nvidia listed a host of early K20 customers, including educational institutions like Clemson University and Indiana University and scientific organizations like the Thomas Jefferson National Accelerator Facility, National Center for Supercomputing Applications and National Oceanic and Atmospheric Administration.
The K20 GPUs are available now with systems from such vendors as Cray, Hewlett-Packard, IBM, SGI and Asus. Users interested in trying the K20 GPU accelerator can do so through remotely hosted clusters via the GPU Test Drive Website.