Next Frontier of Supercomputer: China's Exascale in the Making

[JSCh]

Makers of TaihuLight Supercomputer Offer Commercial Version
Michael Feldman | June 23, 2017 18:33 CEST

One of the more unusual pieces of news at this year’s ISC High Performance conference was the announcement by the National Supercomputing Center in Wuxi that it will be offering a cut-down version of the Sunway TaihuLight supercomputer for more mainstream HPC users.

TaihuLight is the reigning champ on the TOP500 list, delivering a whopping 93 petaflops on the Linpack benchmark. Besides being the number one system, it’s other big claim to fame is that it is constructed almost entirely from Chinese-made componentry. In particular, the system is powered by the 260-core ShenWei processor, known as the SW26010. Each of TaihuLight’s 40,960 ShenWei chips delivers three teraflops of peak performance.

The commercial version they announced at ISC is called the Sunway Micro and is based a dual-socket SW26010 server node. The system is aimed at a broad spectrum of industrial and research applications including “deep learning, oil & gas exploration, climate modeling, etc.”

Source: National Supercomputing Center in Wuxi​

The two-processor design means each node delivers a very respectable six peak teraflops. Unlike the TaihuLight supercomputer, whose single-socket nodes were outfitted with a scant 32 GB of memory, the Sunway Micro can be equipped with 64 GB to 256 GB. That gives Micro buyers the option to have lot more local memory to feed these high-flying ShenWei chips. Each node is also equipped with 12 GB of local storage of undefined type and origin.

While talking with some of the folks at the Wuxi booth during the ISC exhibition, they revealed that the Micro nodes can be clustered together via a network based on InfiniBand technology, which apparently is similar, but not identical to the TaihuLight network implementaion. Given that these servers will be used in relatively small clusters, they didn’t have to develop a network for supercomputer-level scalability.

One of the most unusual aspects of the Sunway Micro is that it is being sold by the National Supercomputing Center in Wuxi. That might seem like an odd thing for a supercomputing center to do, given its public mission. But since the center supplies the system software and developer toolset for these ShenWei-based machines, they basically act as system integrators for the commercial offering. As for the TaihiLight, the Micro was developed by the National Research Center of Parallel Computer Engineering & Technology (NRCPC).

Software support includes C/C++ and Fortran compilers for the ShenWie, as well as supporting runtime libraries. For parallel software development, Wuxi includes MPI, OpenACC and Athread implementations targeted to the ShenWei platform. An integrated development environment, with a debugger and performance monitor, are also included.

Besides selling the standard version of the Micro, the Wuxi center will also provide customized solutions. Pricing for the system was not made public.


Makers of TaihuLight Supercomputer Offer Commercial Version | TOP500 Supercomputer Sites

 

[C130]

https://www.nextplatform.com/2016/06/20/look-inside-chinas-chart-topping-new-supercomputer/

China supercomputers are gimped by inferior software and slow memory. The later is obvious with the SW2060. I wonder how the next gen SW2060 chip will compare to Nvidia V100 and Intel Knights Hill in power,efficiency, and memory bandwidth.

 

[Keel]

• Author: Brian BarrettBrian Barrett
• gear
• 06.21.16
• 03:00 pm

China's New Supercomputer Puts the US Even Further Behind

Jack Dongarra


This week, China's Sunway TaihuLight officially became the fastest supercomputer in the world. The previous champ? Also from China. What used to be an arms race for supercomputing primacy among technological nations has turned into a blowout.

The Sunway TaihuLight is indeed a monster: theoretical peak performance of 125 petaflops, 10,649,600 cores, and 1.31 petabytes of primary memory. That's not just "big." Former Indiana Pacers center Rik Smits is big. This is, like, mountain big. Jupiter big.

TaihuLight's abilities are matched only by the ambition that drove its creation. Fifteen years ago, China claimed zero of the top 500 supercomputers in the world. Today, it not only has more than everyone else—including the United States—but its best machine boasts speeds five times faster than the best the US can muster. And, in a first, it achieves those speeds with purely China-made chips.

Think of TaihuLight, then, not in terms of power but of significance. It’s loaded with it, not only for what it can do, but how it does it.

The Super Supercomputer
If you think of a supercomputer as a souped-up version of what you’re playing EVE Online with at home, well, as it turns out you’re not entirely wrong. “At one level they’re not very different from your desktop system,” says Michael Papka, director of the Argonne Leadership Computing Facility (home to Mira, the world’s sixth-fastest supercomputer). “They have a processor that looks very similar to the one in a laptop or desktop system—there’s just a lot of them connected together.”

Your MacBook, for example, uses four cores; Mira harnesses just under 800,000. It uses those them to simulate and study everything from weather patterns to the origins of the universe. The faster the supercomputer, the more precise the models and simulations.

On that basis alone, TaihuLight is a singular accomplishment. Its 10.6 million cores are more than three times the previous leader, China’s Tianhe-2, and nearly 20 times the fastest U.S. supercomputer, Titan, at Oak Ridge National Laboratory. “It’s running very high rates of execution speed, very good efficiency, and very good power efficiency," says University of Tennessee computer scientist Jack Dongarra. "It’s really quite impressive.”

If anyone’s qualified to say so, it’s Dongarra. He created the benchmark by which supercomputers were first compared in 1993 by TOP500, the organization that still ranks them today, and published the first independent evaluation [PDF] of TaihuLight’s capabilities.

Still, hardware’s not everything. Because supercomputers run specialized tasks, they require specialized software. “You can use a factory as an example,” says Papka. “A lot of people are working on putting a car together at the same time, but they’re all working in a coordinated manner. People who write programs for supercomputers have to get all of the pieces working together.”

Jack Dongarra
TaihuLight passes that test, too. In fact, three of the six finalists for a prestigious high-performance computing award are applications built to run on TaihuLight. Aside from relatively slow memory—a conscious trade off to save money and power consumption—this rig is ready to go to work. “This is not a stunt machine,” says Dongarra. And it's years ahead of anything the US has.

A Command Line Lead
TaihuLight is faster than anything scheduled to come online in the US until 2018, when three Department of Energy sites will each receive a machine expected to range from 150 to 200 petaflops. That’s ahead of where China is now—but two years is half an eternity in computer-time. That the lead has gotten so large galls some lawmakers for reasons both political and practical. Legislation exists calling for a supercomputer funding boost, but has spent the last year mired in the Senate.

“Massive domestic gains in computing power are necessary to address the national security, scientific, and health care challenges of the future,” says Rep. Randy Hultgren, a Republican from Illinois whose American Super Computing Leadership Act has twice been passed by the House of Representatives. “It is increasingly evident that America is losing our lead.” Meanwhile the DOE is working on innovating with the budget it has.

The other significant TaihuLight achievement stings US interests even more, because it's political. China’s last champ, Tianhe-2, had Intel inside. But in February of 2015, the Department of Commerce, citing national security concerns—supercomputers excel at crunching metadata for the NSA and their foreign equivalents—banned the sale of Intel Xeon processor to Chinese supercomputer labs.

Rather than slow the rate of Chinese supercomputer technology, the move appears to have had quite the opposite effect. “I believe the Chinese government put more research funding into the projects to develop and put in place indigenous processors,” Dongarra says. “The result of that, in some sense, is this machine today.”

A Race Worth Winning
Broadly, it's true that better supercomputers benefit the whole world, assuming scientists get to work on them. It doesn't exactly matter what flavor the chips are. “On some level, it’s a trophy that you put on your mantel,” Dongarra says. “But what’s more important is what kind of science it does, what kind of discoveries you make.”

TaihuLight’s stewards tell Dongarra that they're putting all that power toward advanced manufacturing, Earth-system modeling and weather forecasting, life science, and big data analytics. That sounds like a broad range, but it’s just a small slice of what supercomputers' capabilities comprise. “Each time we make an increase, we can add more science to the problem,” Papka says. “For the foreseeable future, until we can model the real world on a quark-for-quark basis, we’ll need more powerful computers.”

And those computers are coming—especially if the US gets serious about catching up.

https://www.wired.com/2016/06/fastest-supercomputer-sunway-taihulight/

 

[Pyr0test]

https://www.nextplatform.com/2016/06/20/look-inside-chinas-chart-topping-new-supercomputer/

China supercomputers are gimped by inferior software and slow memory. The later is obvious with the SW2060. I wonder how the next gen SW2060 chip will compare to Nvidia V100 and Intel Knights Hill in power,efficiency, and memory bandwidth.

memory bottlenecks? Absolutely. Inferior software? Heck no. Not all supercomputing applications are memory intensive, the software written for the machine won Gordon bell prize shows it's amongst one of the best. The article you linked even stated so

 

[Keel]

memory bottlenecks? Absolutely. Inferior software? Heck no. Not all supercomputing applications are memory intensive, the software written for the machine won Gordon bell prize shows it's amongst one of the best. The article you linked even stated so

absolutely
I guess the organisers wont give away the GB Prize for nothing

 

[C130]

memory bottlenecks? Absolutely. Inferior software? Heck no. Not all supercomputing applications are memory intensive, the software written for the machine won Gordon bell prize shows it's amongst one of the best. The article you linked even stated so

I think China has come a long way, but I would bet U.S,EU, and JPN software is still superior.

 

[Keel]

I think China has come a long way, but I would bet U.S,EU, and JPN software is still superior.

Where is the proof?

May be you and your accomplices have to hurry up.
2018 is just a few months away

 

[Han Patriot]

Where is the proof?

May be you and your accomplices have to hurry up.
2018 is just a few months away

Chill guys, he is just trolling.

https://www.nextplatform.com/2016/06/20/look-inside-chinas-chart-topping-new-supercomputer/

China supercomputers are gimped by inferior software and slow memory. The later is obvious with the SW2060. I wonder how the next gen SW2060 chip will compare to Nvidia V100 and Intel Knights Hill in power,efficiency, and memory bandwidth.

Can you explain to me how an inferior software and slower memory managed to beat the fastest American computer?

 

[C130]

Chill guys, he is just trolling.


Can you explain to me how an inferior software and slower memory managed to beat the fastest American computer?

did I ever say it was slower or inferior to anything the U.S has right now?? don't worry Summit,Sierra, and Aurora are coming soon.

Titan is dogcrap right now AMD 16 core+Tesla K20X, woefully outdated. POWER9+Tesla V100 you are looking at 10X performance gain.

 

[Han Patriot]

did I ever say it was slower or inferior to anything the U.S has right now?? don't worry Summit,Sierra, and Aurora are coming soon.

Titan is dogcrap right now AMD 16 core+Tesla K20X, woefully outdated. POWER9+Tesla V100 you are looking at 10X performance gain.

Then what are you comparing with? AIR, genius? Will shall can could, you are starting to sound Indic. Normally, Americans wont say such sh*t, they are quite hands on people.

 

[C130]

Then what are you comparing with? AIR, genius? Will shall can could, you are starting to sound Indic. Normally, Americans wont say such sh*t, they are quite hands on people.

you seem mad. you are taking what I said it having inferior software and slower memory to mean it's slower than U.S and Western supercomputers, haha. It's raw power is undeniable it's 5 times faster than anything the West has right now. calm down son

 

[Han Patriot]

you seem mad. you are taking what I said it having inferior software and slower memory to mean it's slower than U.S and Western supercomputers, haha. It's raw power is undeniable it's 5 times faster than anything the West has, and inferior software/slower memory doesn't change that.

Me mad? you are saying it has inferior software and memory, compared to what? Why do you say it has inferior software? Can you explain to me? Why inferior memory? Please xplain, I want to learn from you. =)

 

[C130]

Me mad? you are saying it has inferior software and memory, compared to what? Why do you say it has inferior software? Can you explain to me? Why inferior memory? Please xplain, I want to learn from you. =)

I already said that U.S,EU, and Japan tend to have better software code for running supercomputers, and I said the SW2060 has gimped memory compared to Knights Landing and Volta, but I also just clarified to my last post that China's Taihu Light has superior performance. not sure why you are getting so butt hurt over me mentioning these two weaknesses that China is or probably has resolved, but U.S and upcoming Japanese HPC will blow away anything China is working on

Tianhe2A is looking like a midget compared to what's coming

https://www.top500.org/news/china-will-deploy-exascale-prototype-this-year/

I was expecting China to deploy a successor to SW2060 with more cores and faster memory, but it looks like it's deploying an ARM and a co-processor which is inferior to POWER9,Volta, and Knights Landing.

 

[Han Patriot]

I already said that U.S,EU, and Japan tend to have better software code for running supercomputers, and I said the SW2060 has gimped memory compared to Knights Landing and Volta, but I also just clarified to my last post that China's Taihu Light has superior performance. not sure why you are getting so butt hurt over me mentioning these two weaknesses that China is or probably has resolved, but U.S and upcoming Japanese HPC will blow away anything China us working on

Tianhe2A is looking like a midget compared to what's coming

https://www.top500.org/news/china-will-deploy-exascale-prototype-this-year/

I was expecting China to deploy a successor to SW2060 with more cores and faster memory, but it looks like it's deploying an ARM and a co-processor which is inferior to POWER9,Volta, and Knights Landing.

Please don't talk to me about what is coming like our Indic friend, you have no idea what the Chinese Peta prototype is right? So stop there buddy. Get back to the topic.

Please tell me why you think US/EU/Japan is having better HPC software, please prove your statement.

What type of memory are you talking about? There is cache memory, there is external memory, please explain, Knights landing is a processor, SW2060 is also a processor.

 

[Beast]

I already said that U.S,EU, and Japan tend to have better software code for running supercomputers, and I said the SW2060 has gimped memory compared to Knights Landing and Volta, but I also just clarified to my last post that China's Taihu Light has superior performance. not sure why you are getting so butt hurt over me mentioning these two weaknesses that China is or probably has resolved, but U.S and upcoming Japanese HPC will blow away anything China is working on

Tianhe2A is looking like a midget compared to what's coming

https://www.top500.org/news/china-will-deploy-exascale-prototype-this-year/

I was expecting China to deploy a successor to SW2060 with more cores and faster memory, but it looks like it's deploying an ARM and a co-processor which is inferior to POWER9,Volta, and Knights Landing.

China has superior software and hardware compare to Europe and US. Statistic dont lie. China supercomputer is the fastest in the world. RIP for western countries. You are no match for China.

Please don't talk to me about what is coming like our Indic friend, you have no idea what the Chinese Peta prototype is right? So stop there buddy. Get back to the topic.

Please tell me why you think US/EU/Japan is having better HPC software, please prove your statement.

What type of memory are you talking about? There is cache memory, there is external memory, please explain, Knights landing is a processor, SW2060 is also a processor.

You know what. When American or Westerner are beaten by hard fact of their system inferior to China one. They will come up with unexplained reason that software is still superior to Chinese one. There is no way to measure software and its very subjective. The only real thing that can be measure is raw power.

Their ego is hurt very badly that is why they need to come up with nonsense to ease their bruise ego. The American and westerner will never admit defeat. They have too much pride to swallow bitter lost.