The 46th edition of the twice-yearly TOP500 list is hot off the presses and while there’s not much to break the monotony at the peak, this is in many ways a pivotal edition of the list in that it makes it hard to dismiss two trends in particular: China’s ascendance and flattening growth trajectories.
But before we unpack those gleanings, let’s start at the top. Yes, Tianhe-2, the supercomputer developed by China’s National University of Defense Technology, is still number one for the sixth time in a row with a LINPACK performance of 33.86 petaflops (54.9 peak); and Titan, a Cray XK7 system installed at DOE’s Oak Ridge National Laboratory, is still number two with 17.59 petaflops LINPACK (27.1 peak).
As a matter of fact, the whole top-five block is right where it was in June 2013:
The only movement the list has seen in the past few years is in the bottom half of the top ten, where we see two newcomers, both Crays. At number six, the Trinity supercomputer, procured by the Alliance for Computing at Extreme Scale (ACES, a Los Alamos and Sandia national laboratories partnership), achieved 8.1 petaflops LINPACK (11 petaflops peak). At number eight, the Hazel-Hen system, installed at the HLRS – Höchstleistungsrechenzentrum Stuttgart, in Germany, reported 5.6 petaflops LINPACK (7.4 peak).
The Cray Trinity install is the first half of a much larger system, reportedly on track to reach 42 peak petaflops when it gets a massive Knight Landing injection in 2016 (it’s aggregate memory capacity will be 2.11 petabytes). The current install is a Cray XC30 based on “Haswells,” specifically Xeon E5-2698v3 16C 2.3GHz processors. When complete, Trinity will be a single system that contains both the Haswell parts and Intel’s next-generation Xeon Phi processors (Knights Landing). The Haswell partition provides a natural transition path for many of the legacy codes running on Cielo, the current NNSA supercomputer sited at Los Alamos. ACES’ RFP partner NERSC is using a similar strategy — and it’s Cori Phase 1 system will be live any day now (read more on NERSC’s new digs here).
“Hazel Hen” at the High Performance Computing Center Stuttgart (HLRS), member of the Gauss Centre for Supercomputing, got on the upper echelon of the list thanks to an upgrade that doubled its compute power from 2.7 to 5.6 petaflops. The Cray XC40 system, now officially open for operation as the most powerful HPC system of PRACE (Partnership for Advanced Computing in Europe), leverages the latest Intel Xeon processor technologies and the CRAY Aries interconnect with Dragonfly network topology. The installation’s 41 system cabinets host 7,712 compute notes with a total of 185,088 Intel Haswell E5-2680 v3 compute cores.
As the TOP500 authors observe, the top ten is experiencing a low level of turnover, a slowing trend that began in 2008. “Six of the top 10 systems were installed in 2011 or 2012, Tianhe-2 in 2013 and only Trinity, Hazel-Hen, and Shaheen II in Saudi Arabia were installed in 2015,” they write.
This stagnation has come to be reflected in long-term performance trajectories with the first, last and sum of systems figures lagging behind historical trends The growth of the average performance of all systems in the list has slowed since 2013, dropping to about 55 percent per years. The performance of the last system on the list (#500) has taken a marked turn in recent years: from 1994 to 2008 it grew by 90 percent per year, but since 2008 it only grows by 55 percent per year, the list authors point out.
Every year of stagnation the flattening effect becomes more clear with a break-point occurring in 2008 and 2013 as Berkeley Lab Deputy Director Horst Simon has elucidated. The numbers would be even flatter if not for the boost provided by strong turnover at the top in the 2011-2013 timeframe. Without a fresh crop of over-achievers, the overall lag has nothing to hide behind. And while there are multi-petaflops systems in the pipeline (the fulfillment of the CORAL and NERSC-8/ACES RFPs, speaking to US interests), they are arriving too late to offset the new trends.
That’s the big picture snapshot, but what this list is even more likely to be remembered for are changing global market dynamics, in which US dominance can no longer be taken for granted. This November, China proved it is serious about filling in its supercomputing portfolio, which it has expanded by nearly a factor of three — going from 37 in July to 109 systems today. This is an unquestionably a steep jump, but a look at China’s system share from year to year shows how this climb has played out since about 2000. The question here of course is whether this ground can be regained or if this is the new reality.
As list share is a zero sum game, the United States has lost TOP500 ground, falling to the lowest point since the TOP500 list was created in 1993 with just 199 systems on this list down from 233 in July. The European share has also declined — down to 107 systems compared to 141 on the last list. Europe now lags behind Asia, which now claims 173 systems, up from 107 from the previous list.
A few more important data points from the TOP500 authors:
- China is at the number two position in terms of both system share and performance share (as it has been since June 2013).
- The European share (107 systems compared to 141 last time) has fallen and is now lower than the Asian share of 173 systems., up from 107 in June 2015.
- Japan has 36 systems, down from 39.
- In Europe, Germany is the clear leader with 32 systems followed by France and the UK at 18 systems each.
- Sugon, a vendor from China is now ahead of IBM in the system category with 49 systems.
There’s still a lot more to unpack. The official list will be published tomorrow and we will follow up with more insights and reporting from the TOP500 BoF, which takes place Tuesday night at the Austin Convention Center.
For now, the TOP500 folks have put together this poster (PDF), which offers a lot to ponder, like chip share, accelerator diversity and the rise of industry.