Supercomputer – Just how fast?
Today’s fastest supercomputer is IBM’s RoadRunner computer housed at the Los Alamos National Laboratory. Running at 1.105 petaflop/second the RoadRunner is narrowly ahead of Oak Ridge National Laboratory’s Cray XT5 supercomputer. That one, named Jaguar, runs at around 1.059 petaflop/s, or more than 1 000 000 000 000 000 floating point operations per second.
In just six years, however, that sort of speed will likely be considered vastly inadequate for “real” computing with experts predicting that the 100 petaflop/second barrier will very likely be broken by 2016.
According to Top500.org, which tracks the world’s largest supercomputers, it has taken just over 20 years for supercomputers to increase performance from Gigaflops to Teraflops to Petaflops. In 1986 Cray2 exceeded the Gigaflops barrier for the first time. Eleven years later the barrier was raised to Teraflops by Intel’s ASCI Red supercomputer in 1997. IBM’s Roadrunner eventually crossed the Petaflops barrier for the first time in 2008, another 11 years after the Teraflops barrier was broken.
The next leap to the Exaflop range is predicted for as early as 2019.
Trends
Hitting these new highs will require significant increases in the numbers of processors underlying supercomputers. Top500.org says that performance improvements will not be achieved by simply increasing per-processor speeds and performance. Instead there will be increasing reliance on clustered multi-core systems to produce the computing power required.
“Going from Terascale to Petascale HPC systems and beyond means that the number of components (cores, interconnect, storage) within such a system will grow enormously. In the near future we will see clustered multi-core systems with core numbers in the range of one hundred thousand to one million and more,” the organisation says.
With increasing numbers of components in systems makers will rely more heavily on parallel computing which itself has a number of challenges. “The shift to the parallel paradigm will be a tough task because there are a lot of programs that have been developed under the premise of a single-core technology. Parallel computer programs are more difficult to write than sequential ones, because concurrency issue will introduce different new classes of potential software bugs.”
Graphics cards
One of the most interesting trends in high performance computing will be the increased appearance of GPU (graphical processing units) in high-end systems. Typically GPUs are used to render high-end graphics on desktop PCs and workstations. But, says Top500.org, as GPUs increase in performance they will become increasingly important to supercomputer makers as they typically offer a better price-performance ratio than most multi-core CPUs.
The one limitation with GPUs is that they are only good for crunching numbers. They are designed to process graphics which means processing data streams and if that could be harnessed alongside CPUs GPUs offer makers attractive options.
The Tsubame supercomputer from the institute of Technology in Tokyo is already proving the value of GPUs and is the first Top500 supercomputer running the Tesla graphics chip from Nvidia. The system-cluster consists of 170 Tesla-S1070-systems running at 170 Teraflops.
Green
The big challenge with supercomputer design in the coming years, however, will be power consumption. With a general trend towards energy-efficient and green computing, high performance computing (HPC) centres will come under increasing pressure to manage power consumption.
Based on current consumption it is estimated that HPC data centres will double in the next four to five years. Most supercomputer centres will be mindful over the coming years of the need to conserve energy and power consumption may actually become the most significant obstacle to future HPC development.
Fastest supercomputer speed discussion
Loading ...