Tesla K20 GPU accelerators ready for supercomputers

The Tesla K20 family of GPUs was recently unveiled by Nvidia Corp. The company's latest GPU accelerators are touted to be its highest performance, most efficient accelerators to date. The massively parallel accelerators are based on the company's CUDA computing platform and programming model.

Launching the new GPUs at the SC12 supercomputing conference in Salt Lake City on Monday (Nov. 12), Nvidia said its flagship Tesla GPU, the K20X, was providing 90 per cent of the performance for the Titan supercomputer at Oak Ridge National Laboratory. That supercomputer has been named the world's fastest on the锟絣atest Top500 list.

Titan boasts 18,688 K20X GPU accelerators along with the same number of 16-core AMD Opteron 6274 processors for a whopping 17.59 petaflops—or 17.59 quadrillion calculations per second.

The Cray XK7 system is made up of no less than 200 cabinets and sports 710 terabytes of memory.

Nvidia's Sumit Gupta, general manager for the Tesla product group, said the use of computers for solving scientific problems hasincreased exponentially—as have energy costs—making the use of accelerators all the more necessary for high-performance computing.

Today, more than 200 software applications take advantage of GPU-acceleration, representing a 60 per cent increase in less than a year.

"When Fermi came along, it was in the right place at the right time,Gupta锟絪aid, predicting that this generation Kepler would see even more adoption in the market.

The new Tesla K20 family features two GPUs, the Tesla K20X and K20. The K20X is said to provide 3.95 teraflops single-precision and 1.31 teraflops double-precision peak floating point performance, while the K20 provides 3.52 teraflops of single-precision and 1.17 teraflops of double-precision peak performance.

Nvidia claims the performance boost is a tenfold application acceleration when paired with leading CPUs.

When paired with Intel Sandy Bridge CPUs, for example, Nvidia said the K20X can accelerate many applications up to 10x or more. That's good news for engineers and scientists alike, with popular programs like Matlab able to run 18.1 times faster, Chroma (for physics) able to run 17.9 times faster, Specfem3D (for earth sciences) able to run 10.5 times faster, and Amber for molecular dynamics running 8.2 times faster.

Better energy efficiency
Nvidia says the new generation is greener too, with the K20X said to deliver three times higher energy efficiency than the previous-generation.

Using Tesla K20X accelerators, Oak Ridge's Titan achieved 2,142.77 megaflops of performance per watt, which surpasses the energy efficiency of the No. 1 system on the most recent Green500 list of the world's most energy-efficient supercomputers.

"We really improved the architectural efficiency," said Gupta. He called beating IBM's Blue Gene Q system in energy efficiency a "big deal."

Customers seem to be buying into the hype, too. At launch, Nvidia could already boast Clemson University, Indiana University, Thomas Jefferson National Accelerator Facility (Jefferson Lab), King Abdullah University of Science and Technology (KAUST), National Centre for Supercomputing Applications (NCSA), National Oceanic and Atmospheric Administration (NOAA), Oak Ridge National Laboratory (ORNL), University of Southern California (USC), and Shanghai Jiao Tong University (SJTU) as being amongst the takers.

The new Tesla's are shipping as of today from manufacturers including Appro, ASUS, Cray, Eurotech, Fujitsu, HP, IBM, Quanta Computer, SGI, Supermicro, T-Platforms and Tyan.

- Sylvie Barak
??EE Times