SoylentNews

takyon writes:

Intel's Knights-branded Xeon Phi chips remain the most familiar "many-core" accelerators or coprocessors. However, another name has emerged recently: PEZY, whose 1,024-core chips were used in the top 3 most efficient supercomputers. Tom's Hardware reports that PEZY's next generation of chips will boost the core count to 4,096 and integrate Imagination's 64-bit MIPS Warrior CPU onto a system-on-a-chip:

PEZY Computing, a Japanese firm that makes the top three most efficient supercomputers in the world, according to the Green500 list, announced that it will integrate Imagination's highly efficient 64-bit I6400 CPUs into its many-core architecture.

The PEZY SC-2 will be PEZY's next-generation system, which will increase the 1024 core count of the first generation PEZY SC to 4096 cores, or four times more. PEZY's many-core accelerator has been combined with Intel CPUs from top supercomputers to significantly increase their efficiency for computing tasks. For instance, the Shoubo supercomputer, which uses Haswell XEON CPUs and PEZY SC many-core accelerators, was able to break the world record with 7 GFLOPS/W performance.

In the November edition of Green500, the top 23 supercomputers used a heterogeneous architecture with many-core accelerators. In the updated June edition of this year, that number increased by 40 percent, and now the top 32 supercomputers are using many-core accelerators. These supercomputers all use accelerators from AMD, Intel, Nvidia and PEZY. The current top 3 supercomputers are manufactured by PEZY Computing and Exascaler Inc, and include Haswell or Ivy Bridge Xeons as well as PEZY many-core accelerators.

Presumably the integration of the MIPS CPU could allow relatively power-hungry Intel Xeons to be ditched entirely.

Previously: MIPS Strikes Back: 64-bit Warrior I6400 Arrives

Original Submission

takyon writes:

HPCwire reports on an analysis of the November 2015 TOP500 Supercomputer list by co-creator Dr. Erich Strohmaier showing "nothing wrong with Moore's Law". Strohmaier examined China's jump in installed systems and performance growth trends.

China's surge is mainly attributed to "surprise company" Sugon, which submitted smaller sytems. It achieved 3rd place in vendor market share, but just 7th in terms of installed performance, with 21 petaflops. Strohmaier says that Sugon was new to supercomputing and took the time and energy to run the LINPACK benchmark across all systems, "regardless of how well or badly they run and gave us the number". Lenovo became a Chinese company, and some "artifact" systems were labelled Lenovo/IBM or IBM/Lenovo. Strohmaier also pointed to Inspur with 15 systems.

Strohmaier identifies two inflection points in TOP500 performance development. The growth trajectory dips in 2008 and 2014, showing the effects of financial and technology changes. Turnover has decreased since 2008, with 1.27 year old systems before 2008 and roughly 3 year old systems today. However, by filtering out systems with NVIDIA and Xeon Phi coprocessors, Strohmaier identified an Rmax/socket trend that continues to follow Moore's Law and is the product of the average number of cores per socket and the performance per core. Since the performance per socket continues to increase at an exponential rate, it is the lack of growth in total number of sockets that explains TOP500 stagnation. "So it's clearly a technological reason, but it's not a reason on a chip, it's actually a reason on the facility and system level that is most likely related to either power or money or both."

[More after the break.]

takyon writes:

The Shoubu supercomputer at RIKEN in Japan continues to lead the Green500 supercomputer efficiency list, but at a lower power efficiency than previously measured now that more processors have been added. Power consumption of Shoubu has tripled from 50.32 kW to 150 kW, and efficiency has declined from 7.03158 gigaflops per Watt to 6.67384 gigaflops per Watt. Say goodbye to that 7 GFLOPS/W milestone for a little while.

Another system at RIKEN, Satsuki, has taken the #2 spot, with 6.19522 GFLOPS/W. Both of these RIKEN supercomputers use Intel Xeon CPUs and PEZY-SCnp "manycore" accelerators. The world's fastest supercomputer, China's Sunway TaihuLight, takes the #3 spot at 6.0513 GFLOPS/W. That supercomputer solely uses a homegrown 260-core processor and consumes a total of 15.371 MW of power.

Despite little movement near the top of the list, there are many new entries this time around:

The Satsuki and TaihuLight supercomputers are the only new entries in the top 10. Overall, there are 157 new systems in the June 2016 edition of the Green500, representing nearly a third of the list. Aside from those systems mentioned, the remaining seven supercomputers in the top 10 use GPUs as accelerators paired with Xeon CPUs. The most energy-efficient systems continue to be dominated by heterogeneous systems like these. In the current list, 40 of the top 50 systems employ some sort of accelerator.

[...] China has 21 of the top 50 greenest supercomputers, while the US claims 8 such systems. Germany has 5 of the top 50 systems, with Japan and France each claiming 4 systems. Looking at the entire list, China has 168 systems, the US has 165, Japan has 29, Germany has 26, and France has 18.

The average energy efficiency in the current list is 1116.8 MFLOPS/Watt or a little over 1 GFLOPS/Watt. While Shoubu, the greenest supercomputer, is more than 6 times as efficient as the average, the goal of a 20 MW exaflop system would require an energy efficiency of 50 GFLOPS/Watt. Using the current trend line, the first 20 MW supercomputer capable of an exaflop would not appear until after 2022.

The TOP500 and Green500 lists have "merged", but the old site is being maintained.

Previously: Shoubu Supercomputer Tops Green500 List at Over 7 Gigaflops Per Watt
TOP500 Analysis Shows "Nothing Wrong with Moore's Law" and the November 2015 Green500 List
TOP500 and Green500 Lists to "Merge"

Original Submission

takyon writes:

The fiftieth TOP500 list has been released. Although there has been little change at the top of the list, China now dominates the list in terms of the number of systems, rising to 202 from 160 in June, with the U.S. falling to 143 systems from 169. However, this seems to be the result of Chinese vendors pushing more commercial systems to get on the list:

An examination of the new systems China is adding to the list indicates concerted efforts by Chinese vendors Inspur, Lenovo, Sugon and more recently Huawei to benchmark loosely coupled Web/cloud systems that strain the definition of HPC. To wit, 68 out of the 96 systems that China introduced onto the latest list utilize 10G networking and none are deployed at research sites. The benchmarking of Internet and telecom systems for Top500 glory is not new. You can see similar fingerprints on the list (current and historical) from HPE and IBM, but China has doubled down. For comparison's sake, the US put 19 new systems on the list and eight of those rely on 10G networking. [...] Snell provided additional perspective: "What we're seeing is a concerted effort to list systems in China, particularly from China-based system vendors. The submission rules allow for what is essentially benchmarking by proxy. If Linpack is run and verified on one system, the result can be assumed for other systems of the same (or greater) configuration, so it's possible to put together concerted efforts to list more systems, whether out of a desire to show apparent market share, or simply for national pride."

Sunway TaihuLight continues to lead the list at just over 93 petaflops. The Gyoukou supercomputer has jumped from #69 (~1.677 petaflops) in the June list to #4 (~19.136 petaflops). Due to its use of PEZY "manycore" processors, Gyoukou is now the supercomputer with the highest number of cores in the list's history (19,860,000). The Trinity supercomputer has been upgraded with Xeon Phi processors, more than tripling the core count and bringing performance to ~14.137 petaflops (#7) from ~8.1 petaflops (#10). Each of the top 10 supercomputers now has a measured LINPACK performance of at least 10 petaflops.

The #100 system has an Rmax of 1.283 petaflops, up from 1.193 petaflops in June. The #500 system has an Rmax of 548.7 teraflops, up from 432.2 teraflops in June. 181 systems have a performance of at least 1 petaflops, up from 138 systems. The combined peformance of the top 500 systems is 845 petaflops, up from 749 petaflops.

Things are a little more interesting on the Green500 list. The Shoubu system B has an efficiency of 17.009 gigaflops per Watt, up from TSUBAME3.0's 14.11 GFLOPS/W at the #1 position in June (TSUBAME3.0 quadrupled its performance while its efficiency dipped to 13.704 GFLOPS/W (#6) on the new list). The top 4 systems all exceed 15 GFLOPS/W. #5 on the Green500 list is Gyoukou, which is #4 on the TOP500. Piz Daint is hanging in there at #10 on the Green500 list and #3 on the TOP500.

All of the new top 3 systems on the Green500 list (and Gyoukou at #5) use the PEZY-SC2 manycore processor. The SC2 has 2,048 cores and 8 threads per core, and has a single-precision peak performance of about 8.192 TFLOPS. Each SC2 also includes six MIPS management cores, making it possible to eliminate the need for an Intel Xeon host processor, although that has not been done in any of the new systems.

At 17 GFLOPS/W, it would take about 58.8 megawatts to power a 1 exaflops supercomputer. 20-25 MW is the preferred power level for initial exascale systems, although we may see a 40 MW system.

Previously: New List of TOP500 Supercomputers [Updated]

Original Submission