The Green500 And The Charge Towards Exascale Computing

The latest Green500 list (Excel spreadsheet here) was just released at the end of June, and heterogeneous systems continue to dominate the top of the list with the new leader besting the old by nearly 30%. Two systems using Intel Xeon E5-2687W 8-Core 3.1 GHz processors coupled with NVIDIA K20’s topped the list. Eurotech’s Eurora and Aurora Tigon both broke the 3,000 MFLOPS/W barrier, recording scores of 3,208.83 MFLOPS/W and 3,179.88 MFLOPS/W respectively, easily outdistancing the past leader at 2.449.57 MFLOPS/W (using Intel Xeon E5-2670 8-Core 2.6 GHz processors coupled with Intel Xeon Phi 5110P’s).

At the moment, the top two computers on the latest Top500 list are also heterogeneous systems. Many experts are expecting this trend to continue. In past articles, we’ve looked at the importance of designing an architecture that’s efficient for the problem at hand. Eckert-Mauchly award winner Bill Dally recently gave the ISC’13 Conference Keynote: Future Challenges of Large-Scale Computing. He presented the slide shown in Figure 1, noting the energy efficiency differences between an architecture optimized for latency (CPU) vs. an architecture optimized for throughput (GPU). It’s this type of energy efficiency difference that has caused heterogeneous computing systems to spring to the top of both the Green500 as well as the Top500 computing lists.

Figure 1. CPU vs. GPU

Dr. Wu-chun Feng, originator of the Green500, presented Figure 2 in the Panel on Exascale Computing, at the GreenGrid Forum in March 2013. It shows how much power would be needed for an ExaFLOPS (10^12 MFLOPS) computer if the efficiency of the computer at the top of the list was extrapolated out to that point. The Green500 line is monotonically decreasing as energy efficiency improves every year, but sometimes a computer can make the top of the Top500 by throwing more power at the problem and it shows a slightly more uneven trend historically. In order to hit a 20 MW target though, we still have a ways to go in terms of better computing efficiency.

Figure 2. Extrapolating Towards Exascale Computing

In Bill’s ISC’13 keynote, he stressed three important points:
• Parallelism is the source of all performance
• Power limits all computing
• Communication dominates power

These are going to be key factors in terms of getting designs to that Exascale goal. The slide in Figure 3 shows Bill’s estimates of where the efficiency improvements will come from in order to reach that goal with architecture playing the most important role.

Figure 3. Improvements Needed to Bridge the Efficiency Gap to Exascale by 2020

Bill also showed that simpler CPU architectures are more energy efficient and systems utilizing ARM cores along with top-end graphic engines are starting to appear in the market. It will be interesting to see how these markets develop and the new innovations that will take place to reach the exascale goal.