China's "Milky Way One" super computer rank fifth in world

[bigmoneymaker]

Well, at least it is clear that the USA still leads in computing power, both in single computer speed and in aggregate national computing capacity. Now, if only we can figure out how to make the Chinese float the yuan .......

yuan floating has anything to do with computing tech??? are you trying to make yuan overevaluating and crumble currently growing chinese economiy?? such disgusting minds of destroying others do exist in the mainstream of usa society, will be a disaster for humanity.

 

[bandmaster]

I think you are mistaken, paritosh.

I am not a computer expert, but I can tell you the following:


1. Currently there are only 2 countries in the world having Petaflop capability, to make its own Petaflop supercomputer. Sorry, none of them is India.


2. The place where a computer is located doesn't count.

e.g. Current # 4 fastest is located in Germany. However, it is not Made-In-Germany,but made-in-USA, with its vendor being IBM. Hence Germany doesn't count in petaflop capability. Otherwise, since #5 China can produce Petaflop, #4 germany surely could count, but that isn't the case.

Get it?

3. The key of supercomputer is the what platform it uses (vendor), and who is the hardware maker, which decides its interllectural ownership.

India's "Hewlett-Packard Cluster Platform 3000 BL460c system"as you mentioned might be rated as #4 sometime in the recent past. However, it only showd that the #4 fastest supercomputer at the time was located in India, nothing less and nothing more. Since its platform (vendor) was HP, and it was Made-In-USA, it intellectural ownershop belongs to USA, not India.

Even though it might have used some Indian techologies in routing, etc, but those are not the core tech of supercomputer, thereby not sufficient enough to claim the ownership of it. The HP supercomputer you have is made-in-USA, only installed in India. It's bit like you bought a Ferrari and made some changes on it at home, but that doesn't mean you produce Ferrari.

On the contrary, the current #5 China's TH-1 is based on Chinese indigenous NUDT platform (vendor), and Made-In-China, which justify its ownership.


I guess Indian media may have misled your domestic audience on this.

You have no expertise and it is showing.
All Top supercomputers are clusters.
The technology to built a cluster is is available to every country where computer hardware is legal.
The cheapest, but not the best, solution is to simply connect a bunch of desktops to a LAN and run MPI software(free).
To get to a practical petaflop, meaning, with as few boxes as possible, you need to use more sophisticated components.
The best I have seen is number 2 on the list. Basically these are Cell processor-based (used in PlayStation 3) multiprocessor blade servers connected by a widely available high-throughput, low-latency technology, Infiniband(Made in Israel).
In fact supercomputers have been built by connecting several PS3's. Search for "Astrophysicist Replaces Supercomputer with a Cluster of Eight PlayStation 3s". Cost 2500 USD .

In short, These are all off-the-shelf components and there is no need to reinvent any wheel.

The NUDT cluster is using off the shelf components too.
Xeon E5540/E5450=>Intel
, ATI Radeon HD 4870 2,=>ATI
Infiniband DDR 4X=>Most likely Mellanox
Operating System=> Linux

The chief motivation is application.
If you need the world's fastest computer, you will build it. As simple as that. Throw me 10000 USD and cost of the system and I will build a petaflop supercomputer for you.

 

[bandmaster]

I'm not too sure but I think it's still based on Intel processor.

China's own processor 4 core Godson-3A successfully loaded with Linux Kernel recently.

That processor does not seem to be meant for a multiprocessing configuration, or it wouldn't have a PCIE unit built in. Most likely, meant for embedded, or desktop.
The processor seems to be MIPS-based so the entire Linux ecosystem can be used without major software modification.

 

[Comet]

Good News. Its great to see China making progress.

 

[lockh33d]

That processor does not seem to be meant for a multiprocessing configuration, or it wouldn't have a PCIE unit built in. Most likely, meant for embedded, or desktop.
The processor seems to be MIPS-based so the entire Linux ecosystem can be used without major software modification.

what does having a PCIE connector have to do with multiprocessing? A PCIE is simply a connector for other peripherals such as the 4870 etc. Please don't post NOOB statements if you don't know what you're talking about. Multiprocessing is to do with the number of physical CPU's and the frequency clocks of each core.

 

[gambit]

If you need the world's fastest computer, you will build it. As simple as that. Throw me 10000 USD and cost of the system and I will build a petaflop supercomputer for you.

Make sure you put the whole she-bang on something that you can claim to be from your native country, like a wood frame or whatever, then you can call it a 'platform' and legitimately claim native design and production.

 

[paritosh]

I think you are mistaken, paritosh.

I am not a computer expert, but I can tell you the following:


1. Currently there are only 2 countries in the world having Petaflop capability, to make its own Petaflop supercomputer. Sorry, none of them is India.


2. The place where a computer is located doesn't count.

e.g. Current # 4 fastest is located in Germany. However, it is not Made-In-Germany,but made-in-USA, with its vendor being IBM. Hence Germany doesn't count in petaflop capability. Otherwise, since #5 China can produce Petaflop, #4 germany surely could count, but that isn't the case.

Get it?

3. The key of supercomputer is the what platform it uses (vendor), and who is the hardware maker, which decides its interllectural ownership.

India's "Hewlett-Packard Cluster Platform 3000 BL460c system"as you mentioned might be rated as #4 sometime in the recent past. However, it only showd that the #4 fastest supercomputer at the time was located in India, nothing less and nothing more. Since its platform (vendor) was HP, and it was Made-In-USA, it intellectural ownershop belongs to USA, not India.

Even though it might have used some Indian techologies in routing, etc, but those are not the core tech of supercomputer, thereby not sufficient enough to claim the ownership of it. The HP supercomputer you have is made-in-USA, only installed in India. It's bit like you bought a Ferrari and made some changes on it at home, but that doesn't mean you produce Ferrari.

On the contrary, the current #5 China's TH-1 is based on Chinese indigenous NUDT platform (vendor), and Made-In-China, which justify its ownership.


I guess Indian media may have misled your domestic audience on this.

you did not understand my post friend.
I never said that we have petaflop capability...the people at C-DAC are working on it.
and as far as the rest of your post...the location is not even a parameter to judge country's supercomputer prowess.
have ytou heard of Cluster Supercomputers?I bet you haven't..cus that is what countries like Germany India and China have been producing all this while...we lack a the tech know-how for producing top-of-the-line processors...the likes of which can compete with those of Intel and AMD...the Americans are way ahead...this milky way SC has American processors...and I believe is a clustered supercomputer....
A supercomputer....even with the availability of off-the-shelf processors dn other parts...is not an easy thing to assemble...or we'd have had a plethora of SCs coming out of Africa....the most important thing to designing a computer of being able to clock in tera and penta flops is the architecture and the architecture alone.

and save the Germany-china argument for later...right now we are all in the same boat...countries that depend solely on ingenious architecture and American processors...
I read some time back that China was developing it's own processors to power it's supercomputers...it has some time.

India's HP Sc was an Indian SC and if you have any doubts regarding that..please read all the news reports that you have posted...it is clearly mentioned that they used
"For the first time ever, India placed a system in the Top 10. The Computational Research Laboratories, a wholly owned subsidiary of Tata Sons Ltd. in Pune, India, installed a Hewlett-Packard Cluster Platform 3000 BL460c system. They integrated this system with their own innovative routing technology and achieved 117.9 TFlop/s performance."

i hope you get what I am trying to say.

 

[bandmaster]

what does having a PCIE connector have to do with multiprocessing? A PCIE is simply a connector for other peripherals such as the 4870 etc. Please don't post NOOB statements if you don't know what you're talking about. Multiprocessing is to do with the number of physical CPU's and the frequency clocks of each core.

I'll ignore the NOOB dig.
Generally a SoC(System On Chip, used in embedded, non-scalable architectures) will have a peripheral controller like PCIe built in.
General purpose CPU's are made to be independent of bus standards, as they keeps evolving. The latest is PCIE2 and PCIE3 is already in trial, for example.
The general way to make a a server class CPU, is to have a highspeed north bridge. The bridge, which will be part of chipset( not the processor) will handle the peripheral protocol.

Why I believe it is not multiprocessor? Because 2 PCIE units per CPU is not needed. An 8-way SMP will have 16 of these. There is no correlation between CPU scaling and I/O attachment requirement.

 

[Merilion]

That processor does not seem to be meant for a multiprocessing configuration, or it wouldn't have a PCIE unit built in. Most likely, meant for embedded, or desktop.
The processor seems to be MIPS-based so the entire Linux ecosystem can be used without major software modification.

Yes, the processor is MIPS-based.
Dawning is building its next generation supercomputer Dawning6000 based on Godson-3 processors. The Dawning6000 will have 8000 Godson-3 processors and will be in service in year 2010.
Dawning current supercomputer Dawning5000A is ranked #19 in the 100 supercomputer list. It's using 8000 AMD processors. Dawning6000 will be 4x faster than Dawning5000A.
ÎÒ¹úǧÍòÒڴμÆËã»ú½«ÓðËǧ¿ÅÁúо3ºÅCPU-¼ÆËã»ú,Áúо3ºÅ,CPU-µç×ÓÐÐÒµ-hc360»Û´ÏÍø a chinese url for those who can read chinese.

PS: this is just for information sharing, it's not a size contest.

 

[bandmaster]

Yes, the processor is MIPS-based.
Dawning is building its next generation supercomputer Dawning6000 based on Godson-3 processors. The Dawning6000 will have 8000 Godson-3 processors and will be in service in year 2010.
Dawning current supercomputer Dawning5000A is ranked #19 in the 100 supercomputer list. It's using 8000 AMD processors. Dawning6000 will be 4x faster than Dawning5000A.
ÎÒ¹úǧÍòÒڴμÆËã»ú½«ÓðËǧ¿ÅÁúо3ºÅCPU-¼ÆËã»ú,Áúо3ºÅ,CPU-µç×ÓÐÐÒµ-hc360»Û´ÏÍø a chinese url for those who can read chinese.

PS: this is just for information sharing, it's not a size contest.

Thanks for taking the discussion "forward" and not degrading it.
The Godson-3 appears to be a general reconfigurable architecture.
The question is about the particular diagram posted above.
In fact the diagram is incorrect . All papers I've seen show that the PCIE is connected to XBAR2 at the bottom through a switch/bridge.
The focus seems to be on multicore scaling as opposed to multiprocessor SMP scaling.

 

[lockh33d]

I'll ignore the NOOB dig.
Generally a SoC(System On Chip, used in embedded, non-scalable architectures) will have a peripheral controller like PCIe built in.
General purpose CPU's are made to be independent of bus standards, as they keeps evolving. The latest is PCIE2 and PCIE3 is already in trial, for example.
The general way to make a a server class CPU, is to have a highspeed north bridge. The bridge, which will be part of chipset( not the processor) will handle the peripheral protocol.

Why I believe it is not multiprocessor? Because 2 PCIE units per CPU is not needed. An 8-way SMP will have 16 of these. There is no correlation between CPU scaling and I/O attachment requirement.

Exactly, "There is no correlation between CPU scaling and I/O attachment requirement." So what is the relation between the CPU and the number of PCIE bridges in correlation to the multi-processing ability of the computer as a whole?

 

[Speeder]

You have no expertise and it is showing.
All Top supercomputers are clusters.
The technology to built a cluster is is available to every country where computer hardware is legal.
The cheapest, but not the best, solution is to simply connect a bunch of desktops to a LAN and run MPI software(free).
To get to a practical petaflop, meaning, with as few boxes as possible, you need to use more sophisticated components.
The best I have seen is number 2 on the list. Basically these are Cell processor-based (used in PlayStation 3) multiprocessor blade servers connected by a widely available high-throughput, low-latency technology, Infiniband(Made in Israel).
In fact supercomputers have been built by connecting several PS3's. Search for "Astrophysicist Replaces Supercomputer with a Cluster of Eight PlayStation 3s". Cost 2500 USD .

In short, These are all off-the-shelf components and there is no need to reinvent any wheel.

The NUDT cluster is using off the shelf components too.
Xeon E5540/E5450=>Intel
, ATI Radeon HD 4870 2,=>ATI
Infiniband DDR 4X=>Most likely Mellanox
Operating System=> Linux

The chief motivation is application.
If you need the world's fastest computer, you will build it. As simple as that. Throw me 10000 USD and cost of the system and I will build a petaflop supercomputer for you.

I said at the beginning that I am not an expert on computer. However, I am an expert on logic.

Your post didn't debunk my previous one, cuz you have several severe logic problems which are easy to spot.

1. supercomputer may use off-the-shelf third-party components. It seems however, that this is not amongst criteria which decide the ownership of it. Otherwise, China's TH-1 would not be termed as China's own interllectural property. Of course TH-1 uses some third-party processors. And? What's your point?

2. vast evidences show that supercomputer is not simply the result of money, as you almost idotically claimed. Some simple supercomputer might be. The more one approaches to the top rank, the more technologies needed other than just money.

Otherwise any nation can at least afford to build one petaflop itself , yet in fact there are only 2 countries in the f***ing world can build petaflop supercomputer, while you said " Throw me 10000 USD and cost of the system and I will build a petaflop supercomputer for you". What, are you a Martian?

3. It seems to me that strong evidences pointing to areas of system/platform design - in a sense of how to intergrate the whole stuffs together, and quality of components used are 2 key areas to determine the overall performance, with system/platform design seems to be the most important part. Both of these 2 criteria require hell lots of high tech, which most countries in the world don't qualify.

Again, use car as an analogy:

Most cars use some critical third party components, such as high-performance Bridgestone tyres and specially-made steel products, ect. However, Bridgestone and steel companies can't claim ownership of all the car brands.

A car's engine undountedly is the most important part of the car. Most car engines , including Ferrari, still use some third party components. However, as I said system/platform design - how to design the high performance engine to put piece together, and quality of components used decide engine performance (particularly system/platform design), and determine which is the one to make it a Ferrari, and which is the one to make it a joking Tata Nano.

China's TH-1 is using China's own NUDT platform ( "engine design"), not HP, IBM, Sun, Clay, etc which are alternative platforms used elsewhere. This mainly makes the whole difference.



Now make sense, " pundit " ?


Paritosh, my above comment also serves as the response to your most recent remark.

A computer-major expert is needed in this board , now!

 

[grey boy 2]

China's fastest supercomputer unveiled- China - Guangdong

China's fastest supercomputer unveiled
Latest Updated at 2009-October-30 09:42:37
Story Highlights

* The supercomputer is theoretically able to do more than 1 quadrillion calculations per second (one petaflop) at peak speed.

* Tianhe was able to store all 27 million books in the National Library of China four times over.

* The performance of Tianhe would have made it the world's fourth most powerful supercomputer in the most recent ranking in June.

The National University of Defense Technology (NUDT) unveiled Thursday (Oct 29) China's fastest supercomputer, which could rival the world's most powerful computing devices.

The supercomputer, named Tianhe, meaning Milky Way, is theoretically able to do more than 1 quadrillion calculations per second (one petaflop) at peak speed.

A single-day task for Tianhe might take a mainstream dual-core personal computer 160 years to complete, working non-top -- if it can last that long.

NUDT president Zhang Yulin said the 155-ton system, with 103 refrigerator-like cabinets lined up on an area of about 1,000 square meters, is expected to process seismic data for oil exploration, conduct bio-medical computing and help design aerospace vehicles.

China's national high-technology research and development program and the Binhai New Area, a major economic development zonein the northern port city of Tianjin jointly financed Tianhe, which cost at least 600 million RMB (88.24 million USD).

China's fastest super computer "Tianhe", meaning Milky Way, is seen in Changsha, capital of central China's Hunan Province, Oct. 29, 2009. The National University of Defense Technology (NUDT) unveiled Thursday China's fastest supercomputer, which could rival the world's most powerful computing devices. The supercomputer is theoretically able to do more than 1 quadrillion calculations per second (one petaflop) at peak speed.(Xinhua/He Shuyuan)

Tianhe's peak performance reaches 1.206 petaflops, and it runs at 563.1 teraflops (1,000 teraflops equals one petaflop) on the Linpack benchmark, which was originally developed by U.S. computer scientist Jack Dongarra and has become an internationally recognized method to measure a supercomputer's real performance in practical use.

Zhang said the technical data of Tianhe had been submitted to the world Top 500 list, compiled by the University of Mannheim, in Germany, the National Energy Research Scientific Computing Center, the Lawrence Berkeley National Laboratory, and the University of Tennessee in the United States.

The next Top 500 supercomputer list will be released in November.

The performance of Tianhe would have made it the world's fourth most powerful supercomputer in the most recent ranking in June.

"I was shocked at the milestone breakthrough, which was beyond expectation," said Zhang Yunquan, a researcher with the Institute of Software of the Chinese Academy of Sciences (CAS), and an organizer of the China Top 100 list, which was released at a national conference on high-performance computers Thursday.

The National University of Defense Technology (NUDT) unveiled Thursday China's fastest supercomputer, which could rival the world's most powerful computing devices. The supercomputer is theoretically able to do more than 1 quadrillion calculations per second (one petaflop) at peak speed.(Xinhua)

"I previously forecast China's first petaflop computer no earlier than the end of 2010," Zhang said.

The giant device, a product of 200 computer scientists and two years' work, was housed in the NUDT campus in Changsha, Hunan Province, and would be moved to the National Supercomputing Centerin Tianjin at the end of 2009, said Li Nan, chief coordinator of the program.

Equipped with 6,144 Intel CPUs and 5,120 AMD GPUs, Tianhe was able to store all 27 million books in the National Library of China four times over, said Zhou Xingming, an academician of CAS and a professor with NUDT.

"As far as I know, a combination of CPU and GPU is something new used to make a petaflop computer. A GPU, or graphic processing unit, plays a role as an accelerator to make the computer run faster, but reduces its power consumption and cost," Zhou explained.

"After it's installed in Tianjin, we plan to add hundreds or thousands of China-made CPUs to the machine, and improve its Linpack performance to over 800 teraflops," Zhou said.

The National University of Defense Technology (NUDT) unveiled Thursday China's fastest supercomputer, which could rival the world's most powerful computing devices. The supercomputer is theoretically able to do more than 1 quadrillion calculations per second (one petaflop) at peak speed.(Xinhua)

Although its annual electricity bill can be as high as 18 million RMB, Tianhe could have been ranked the world's fifth greenest supercomputer, according to Green 500 List in June, compiled by researchers at Virginia Tech aiming to provide a ranking of the most energy-efficient supercomputers in the world and serve as a complementary view to the Top 500.

Of the world's fastest 500 supercomputers, the United States alone has invented 291, including the top 10, Europe has 145 and Asia 49, the June World Top 500 List said.

In the same list, the Chinese mainland has 20 high-performance computers, with CPUs all supplied by foreign manufacturers.

China's Dawning Information Industry Company is attempting to build its own supercomputer that overcomes the petaflop barrier by 2010.

 

[Speeder]

China's fastest supercomputer unveiled- China - Guangdong

The next Top 500 supercomputer list will be released in November.

The performance of Tianhe would have made it the world's fourth most powerful supercomputer in the most recent ranking in June.

"I was shocked at the milestone breakthrough, which was beyond expectation," said Zhang Yunquan, a researcher..

The National University of Defense Technology (NUDT) unveiled Thursday China's fastest supercomputer, which could rival the world's most powerful computing devices. The supercomputer is theoretically able to do more than 1 quadrillion calculations per second (one petaflop) at peak speed.(Xinhua)

"I previously forecast China's first petaflop computer no earlier than the end of 2010," Zhang said.

The giant device, a product of 200 computer scientists and two years' work, was housed in the NUDT campus in Changsha, Hunan Province, and ...

Equipped with 6,144 Intel CPUs and 5,120 AMD GPUs, Tianhe was able to store all 27 million books in the National Library of China four times over, said Zhou Xingming, an academician of CAS and a professor with NUDT.

"As far as I know, a combination of CPU and GPU is something new used to make a petaflop computer.. A GPU, or graphic processing unit, plays a role as an accelerator to make the computer run faster, but reduces its power consumption and cost," Zhou explained

This has confirmed my initial thoughts: design is the most critical thing !

 

[bandmaster]

I said at the beginning that I am not an expert on computer. However, I am an expert on logic.

Your post didn't debunk my previous one, cuz you have several severe logic problems which are easy to spot.

1. supercomputer may use off-the-shelf third-party components. It seems however, that this is not amongst criteria which decide the ownership of it. Otherwise, China's TH-1 would not be termed as China's own interllectural property. Of course TH-1 uses some third-party processors. And? What's your point?

2. vast evidences show that supercomputer is not simply the result of money, as you almost idotically claimed. Some simple supercomputer might be. The more one approaches to the top rank, the more technologies needed other than just money.

Otherwise any nation can at least afford to build one petaflop itself , yet in fact there are only 2 countries in the f***ing world can build petaflop supercomputer, while you said " Throw me 10000 USD and cost of the system and I will build a petaflop supercomputer for you". What, are you a Martian?

3. It seems to me that strong evidences pointing to areas of system/platform design - in a sense of how to intergrate the whole stuffs together, and quality of components used are 2 key areas to determine the overall performance, with system/platform design seems to be the most important part. Both of these 2 criteria require hell lots of high tech, which most countries in the world don't qualify.

Again, use car as an analogy:

Most cars use some critical third party components, such as high-performance Bridgestone tyres and specially-made steel products, ect. However, Bridgestone and steel companies can't claim ownership of all the car brands.

A car's engine undountedly is the most important part of the car. Most car engines , including Ferrari, still use some third party components. However, as I said system/platform design - how to design the high performance engine to put piece together, and quality of components used decide engine performance (particularly system/platform design), and determine which is the one to make it a Ferrari, and which is the one to make it a joking Tata Nano.

China's TH-1 is using China's own NUDT platform ( "engine design"), not HP, IBM, Sun, Clay, etc which are alternative platforms used elsewhere. This mainly makes the whole difference.



Now make sense, " pundit " ?


Paritosh, my above comment also serves as the response to your most recent remark.

A computer-major expert is needed in this board , now!

A petaflop is simply a measure of floating point operations per second. If I buy a 1000 1 teraflop systems(see Nvidia's GPU based desktop supercomputer) and network them, I will have a theoretical petaflop.
For running FFT algos the petaflop performance will not be realized, because of high communication fanout of FFT.
For some embarrassingly parallel algos like folding@home, something close to theoretical peak can be achieved, even over a slow communication network.
In fact, folding@home actually was the first "cluster" to achieve a petaflop. You can run the client on your system right now, and you would have added a few gigaflops to it.
I never said I would give you a petaflop for every conceivable algorithm or that it would qualify for Top500.
I am merely proposing to sell thousands of interconnected CPU/GPUs with a 10000 USD premium.

The number 2 system on the list, Roadrunner, is a purely COTS system. You can buy all the components from the market and build the exact same supercomputer.