June 17, 2013
Moments ago the founders of the Top 500 supercomputer rankings unleashed the results of their 41st incarnation of the list at the International Supercomputing Conference in Leipzig, Germany.
And unlike other years where heated speculation about the victors swirled before the show, this announcement of the winner came as little surprise. Word about the Xeon Phi-boosted Chinese Tianhe-2 system leaked, but was quickly validated by Dr. Jack Dongarra after his own visit to China in the last couple of weeks.
The initial reports rolled down the mountain following a gun-jumping presentation by the Chinese during the International HPC Forum in Changsha, China in late May during which all was (prematurely) revealed--in part, if one might speculate, to show the supercomputing dons that the fabled next-gen Milky Way was tangible, operational and ready to rock collective Linpack socks.
The real story during this iteration of benchmarks deviates from the dead heat races of highly similar top-end systems vying for slivers of performance differentiation. In fact, there hasn’t been much change at all for the usual suspects that haunted November’s rankings. Even the mighty Titan’s numbers are the same since the team decided to stable their race horse and use November results.
Some systems, including Sequoia shined a bit brighter with a fresh run of the benchmark. And while Titan was kicked to the number two slot, it still keeps its crown as one of the most efficient systems at the top, consuming 8.21 MW and hitting the 2.143 MFlops/W, just outpacing Sequoia’s 7.84 MW consumption, which packs 2,031.6 MFlops/W.
But Tianhe-2, with its 3 million processors, early-gen MIC cards and unique Chinese-crafted stacks from interconnect to inner software workings is what is sparking the conversational fuse today. And it comes as no surprise, because--and let’s not mince words here--it blew everything out of the freakin’ water.
To put the above image from the Top500 presentation this morning in a bit more context, take a look at the side by side comparison of the two architectures in the chart below, paying specific attention to the power and cooling numbers in particular. It’s not just that these are both relatively efficient systems for their size and our current expectations--it’s that the Chinese are taking a giant leap toward exascale, and are finding ways to stuff the power envelope.
And so what does this upset to the standard order of close top contender races really mean? The significance reaches across the themes of accelerators and co-processors as a complex, if not contradictory trend. It pushes new ideas about how to shove the steady curve of Moore’s law under the 25 MW umbrella...And of course, it does a fair job of shaking the global supercomputing technology market’s westward-focused worldview.
To the last point, As Dongarra noted, “Most of the features of the system were developed in China and they are only using Intel for the main compute part. That is, the interconnect, operating system, front-end processors and software tools are mainly Chinese in origin.”
While the system offers some notable publicity for Intel, which still owns an 80.4% processor share across the Top 500 (of around 800 or so submitted systems), some see this as the dawn of a new era for foreign components and systems. While Tianhe-2 is not Godson based (which was the initial inkling before details leaked), it’s only a matter of time before China puts natively-built processor, accelerator and coprocessor capstones on its masterpieces.
It’s not clear what the next big super out of the United States, the anticipated Trinity system at Los Alamos National Lab, will flop in at. But in case the world missed the message with the first Tianhe machine, China has signaled its entry into the supercomputing spectrum in earnest. As a nation, this June ranking reveals that China now is the second largest user of HPC, outpacing Japan, the UK, France and Germany--all supercomputer strongholds.
And then what? Most likely some new interconnect concepts matched with memory innovations that strike efficiency concerns lower to the ground. What we have then is a global supercomputing race that puts some bold new players into the game. If it weren’t for missing out on summer, it’s tempting to want to rush the clock forward to November just to see what happens from this tiny spark.
Top 500 founder Hans Meuer equated the announcement of the system as being as disruptive as the Earth Simulator, which was 11 years ago. It bumped off its rivals by 5x and set the stage for a new era...one that he says we’re entering, even if it’s sketchy what will smooth the curves of Moore’s Law.
As Horst Simon told us, if the total power consumption is 25 MW (including cooling) we’d all like to see systems operate at that kind of power envelope but there is still the need to produce more energy efficient technology. “We are going on a Moore’s law curve; the areas where improvement is needed is on interconnects and memory.”
The problem is, the founding fathers we spoke with yesterday can’t see what lies beyond the bend of the curve. Simon noted, “Unlike in the past where we could see three or four years in the future, at this point, I don’t see anything radically different coming up. We have reaped all the benefits from the standard architectures to manycore--the magic bullet has already been shot.”
That magic bullet, of course, is the mighty accelerator.
Analysts agree across the board that accelerators will be the kings of the top supers--beyond the top ten in a few years. As Dongarra told us, however, if you look at the list, the number of machines with accelerators has gone down. However, if you look at the sum total of the 500 systems in terms of overall power, it’s 33 percent aggregate performance, mostly due to the big DoE machines that occupy the limelight.
Dongarra highlighted that it’s 33 percent aggregate with this list, but on the last list it was 22% and before that, just 11%.
Horst Simon explained that as of now, accelerators aren’t being used to much of an extent in the commercial systems they’ve seen submitted. There is a bridging of gaps that needs to happen on the learning curve side but he says that the right tools for big businesses on big iron (GPUs matched with powerful database tools, for instance) will bring others on board with acceleration, evening out the list.
By the end of the year the Tianhe-2 super is fully operational at its temporary home at the National University for Defense Technology in Changsha, China. Before the year is out it will find its way to its permanent spot at the National Supercomputer Center in Guangzho and is trained to crunch a number of common supercomputing applications in the fields of biomedical research, climatology, manufacturing and beyond
Even with Trinity rising on the 2015-ish horizon, when asked how long Tianhe-2 will stand at the top, Dongarra laughed softly and said, “a long time...probably a pretty long time.”
10/30/2013 | Cray, DDN, Mellanox, NetApp, ScaleMP, Supermicro, Xyratex | Creating data is easy… the challenge is getting it to the right place to make use of it. This paper discusses fresh solutions that can directly increase I/O efficiency, and the applications of these solutions to current, and new technology infrastructures.
10/01/2013 | IBM | A new trend is developing in the HPC space that is also affecting enterprise computing productivity with the arrival of “ultra-dense” hyper-scale servers.
Ken Claffey, SVP and General Manager at Xyratex, presents ClusterStor at the Vendor Showdown at ISC13 in Leipzig, Germany.
Join HPCwire Editor Nicole Hemsoth and Dr. David Bader from Georgia Tech as they take center stage on opening night at Atlanta's first Big Data Kick Off Week, filmed in front of a live audience. Nicole and David look at the evolution of HPC, today's big data challenges, discuss real world solutions, and reveal their predictions. Exactly what does the future holds for HPC?