Students Use GPUs to Sweep Asian Supercomputing Competition, Top Green500

by Roy Kim

After a grueling five-day battle, three teams emerged as winners in the first annual Asia Student Cluster Challenge (ACS13).  One even beat the power-efficiency record set by the current Green500 leader. Their secret sauce: NVIDIA Tesla K20 accelerators.

ACS13 brought the brightest young minds from ten universities, from Saudi Arabia to South Korea, to out-smart and out-design each other to construct the fastest, most energy efficient supercomputer, in less than a week.  And they had to build a supercomputer that used no more than 3000 watts – about the power of 3 microwave ovens.

Teams were ranked on how their cluster performed on five benchmark applications, such as the GROMACS molecular simulation software and the OpenCFD aerodynamics simulation software, as well as Linpack, which is an industry standard benchmark to measure the performance of supercomputers.

Since it was an Intel-sponsored event, all the teams were also required to use Xeon Phi accelerators to run an application based on the BSDE options pricing algorithm used in the financial services industry.

The three winning teams stood above the rest in the end because they all used NVIDIA GPU accelerators.  Tsinghua University in Beijing took first place, Taiwan’s National Tsing Hua University was second and Huazhong University of Science and Technology took the third spot.

The final results.

The winning team from China’s Tsinghua University built a 4 server node cluster, each server node with 2 Intel Xeon CPUs and  2 NVIDIA Tesla K20 accelerators, to deliver the highest Linpack score of 7.6 teraflops – or trillion floating point operations per second.  They also maximized their GPU usage for other applications, like OpenCFD and GROMACS, to rise to the top.

If that’s not impressive enough, their cluster edged out the most energy efficient system on the current Green500 list, delivering 2526 megaflops per watt compared to the reigning Green500 champ, Beacon from the University of Tennessee’s National Institute for Computational Sciences, at 2499 megaflops per watt (a megaflop is equal to one million floating-point operations per second).

The winning team.

Not too shabby for a few university students in a few days of work.

Taiwan’s National Tsing Hua University and Huazhong University of Science and Technology also used Tesla K20 accelerators to achieve impressive benchmarks on their way to the podium for silver and bronze medals.

Thanks to the sponsors of the ACS13 for a wonderful event and congratulations to the three winning teams.