Green Computing: GPUs Strike Back

Speed matters, but so does efficiency.

popularity

After the last Platform for Advanced Scientific Computing Conference in June, I wrote an article here about how the custom designed chips from NRCPC (used in the Sunway TaihuLight) and PEZY Computing (used in the PEZY Shoubu) had jumped to the top of the Green500 list with the Sunway TaihuLight also remarkably topping the Top500 list. Well, six months after the report from the IEEE/ACM SC16 Conference, the Sunway TaihuLight still tops the Top500 list as the world’s fastest supercomputer, but has slipped to 4th on the Green500 list and the PEZY Shoubu is now in 3rd on the Green500 list with two systems (DGX SaturnV and Piz Daint) both based on NVIDIA Tesla P100s at the #1 and #2 positions on the Green500. The new Tesla P100s also rank 28th and 8th respectively on the Top500 list. It should also be noted that Piz Daint has appeared previously on the Green500 list, but has been recently updated with new NVIDIA Tesla P100s replacing its older NVIDIA K20x processors that were originally installed in the system. Figure 1 below shows the Green100 systems plotted on a log-log chart.

efficiency_vs_performance
Figure 1. Efficiency vs. Performance

The new Tesla P100 based systems offer a significant improvement in energy efficiency over the previous #1 on the Green500 with the SaturnV coming in just under 10 GFLOPS/W at 9.462 GFLOPS/W and the Piz Daint at 7.453 GLOPS/W (the old Piz Daint was at 2.697 GLOPS/W and Rmax of 6,271 TeraFLOPS). The PEZY system comes in at 6.674 GFLOPs/W so the two new systems are 41.8% and 11.7% more energy efficient in terms of FLOPS/W. This also tends to restore the “natural order” with a line through the highest and rightmost systems shown in Figure 1 now having a more negative slope. With a Grand Challenge goal of 1 ExaFLOPS/20MW (equivalent to 50 GFLOPS/W), these new systems with their current leap in efficiency are still a factor of 5.28x and 6.71x away from an efficiency standpoint and a factor of 302x and 102x away in terms of performance. It appears that there’s still plenty of work to do to reach the Grand Challenge Exascale goal.

Sandwiching the other side of the Sunway and PEZY systems are three new Intel Xeon Phi 7200 series based systems; the QPACE3, Oakforest-PACS and the Theta with the first two manufactured by Fujitsu and the third by Cray. These 3 systems have efficiency ratings of 5.806 GFLOPS/W, 4.986 GLOPS/W and 4.688 GFLOPS/W and rank 375th, 6th and 18th on the Top500.

Intel makes a strong showing in the first 10 systems on the Green500 list with the Sunway being the only machine not to use Intel silicon. There are 5 new Intel Xeon Phi 7200 series based machines in the 10 most energy efficient machines and 4 others, including the top 3, are using Intel Xeon E5’s.

Since the Xeon Phi’s and Tesla P100’s are recently introduced, there isn’t expected to be another jump here until late 2017 or early 2018 when the next generation of these products are released, unless one of the other manycore processor manufacturers, like China’s Sunway or Japan’s PEZY, release something new sooner.

Related Stories
GPUs Power Ahead
AI, ADAS, deep learning, gaming, and high performance computing continue to drive GPUs into new application areas.
CPU, GPU, Or FPGA?
Need a low-power device design? What type of processor should you choose?
How To Choose A Processor
There is no clear formula for what to use where, but there are plenty of opinions about who should make the decision.



  • realjjj

    “there isn’t expected to be another jump here until late 2017 or early ,2018”

    AMD Vega 10 should arrive before that and the first product announced with it , the Radeon Instinct MI25, seems to be advertised at some 25TFLOP FP16 at under 300W. Would beat P100 ,not by a huge margin but relevant.