The group that measures the world's Top 500 fastest supercomputers hasn't crowned a new champion in more than a year.
Tianhe-2, of China’s National Super Computer Center, took over the top spot in June 2013 with a measured speed of 33.86 petaflop/s, and it held on to #1 in both the November 2013 list and the June 2014 list released yesterday.
The follow-up to Tianhe-1A, Tianhe-2 uses Ivy Bridge-based Intel Xeons and Intel Xeon Phi for a total of 3.12 million cores. The computer uses 17,808 kilowatts of power for 1.9 gigaflop/s per watt and can theoretically hit speeds of up to 54.9 petaflops.
A system maintaining the #1 spot three times in a row isn't unprecedented: IBM's US-based Roadrunner, the first petaflop machine, won three straight titles from June 2008 to June 2009. But in the latest list, the top nine machines are identical to those from six months ago. And even the bottom of the Top 500 is seeing less turnover and growth than usual.
"Since its inception in June 1993, the TOP500 list has served as a consistent measure of the performance growth of supercomputers, since all systems are ranked according to performance running the same Linpack benchmark application," yesterday's announcement said. "For the second consecutive list, the overall growth rate of all the systems is at a historical low."
The only new entry in the top 10 "was at number 10—a 3.14 petaflop/s Cray XC30 installed at an undisclosed US government site," the Top 500 project leaders wrote. A petaflop is one quadrillion, or a thousand trillion, calculations per second.
“Things seem to be slowing down,” University of Tennessee professor Jack Dongarra, who created the Linpack benchmarks and helps compile the bi-annual Top 500 list, told us “You might characterize it as maybe a sign that Moore’s Law is having some issues.”
Some further stats help illustrate the slowdown in growth. This time around, the very last system on the Top 500 list was previously rated as the 384th fastest system in November 2013. "This represents the lowest turnover rate in the list in two decades," the Top 500 announcement said. In other words, fewer new systems are joining the Top 500: In June 2013, the 500th system had fallen all the way from #322.
Between 1994 and 2008, the very last system on the Top 500 list grew performance by an average of 90 percent each year. Since then, the #500 system has improved its performance just 55 percent each year.
The combined performance of all 500 systems hit 274 petaflop/s in the latest list, up from 250 petaflop/s six months ago and 223 petaflop/s one year ago. "This increase in installed performance also exhibits a noticeable slowdown in growth compared to the previous long-term trend," the announcement said.
Supercomputer makers have been boosting speed in part by using co-processors as "accelerators" to handle some of the work that would otherwise be done by CPUs. Sixty-two of the Top 500 systems have co-processors, with 44 of those using Nvidia's graphics processing units. Nvidia is hoping to make ARM a major part of the supercomputing world by pairing 64-bit ARM server processors with its GPU accelerators.
Vendors could increase supercomputer performance a lot more simply by creating ever bigger systems—but it wouldn't necessarily be efficient. Companies are trying to figure out how to efficiently create exascale supercomputers, which would be 1,000 times faster than a petaflop per second. Intel today announced new Xeon Phi processors and a more efficient and lower-latency interconnect, calling the new architecture "the first viable step towards exascale."
Two years ago, Intel said 40 to 50 gigaflops of performance per watt is needed to hit an exaflop, but that milestone is thought to be at least several years away. The most efficient Top 500 supercomputer, which is at the Tokyo Institute of Technology and uses both Intel CPUs and Nvidia GPUs, can hit 4.5 gigaflops per watt.
No comments:
Post a Comment