China unveils its fastest supercomputer-Tianhe

RAM

The southern Man
Senior Member
Joined
Jul 15, 2009
Messages
2,288
Likes
445
Country flag
China unveils its fastest supercomputer-Tianhe

New Delhi: China has unveiled its fastest supercomputer, which at its peak speed can do more than one quadrillion calculations per second, known as a petaflop.
The birth of the supercomputer, named 'Tianhe', which means 'Milky Way', makes China the second country, after the United States, to build a petaflop computer.
This gigantic device can do as many calculations in one day as those done by an ordinary dual-core personal computer in 160 years, according to Li Nan, director of the Tianhe Project Office under the National University of Defense Technology, developer of the supercomputer.

Tianhe, at a cost of 600 million yuan, ranks first on China's Top 100 supercomputer list released in Changsha, capital of Hunan Province, with a theoretical peak performance of 1.206 petaflops and a Linpack performance of 560 teraflops.

The Linpack benchmark is used to measure the supercomputer's real performance in practical use.

The data has been submitted to the world Top 500 organisation, which will release the new ranking list in November.

According to the Top 500 list in June, Tianhe can be ranked the fourth most powerful supercomputer in the world.

"Supported by the national high-tech research and development programme, also known as the 863 programme, development of Tianhe has involved more than 200 computer experts who spent about two years in designing and producing the petaflop supercomputer," said Zhang Yulin, president of the defense university.

The supercomputer has many practical applications, and will be used in seismic data processing for oil exploration, bio-medical research, development of aerospace vehicles, long-term weather and climate forecasting, financial data analysis and pollution control in the Bohai Sea area, according to Li.

Tianhe, made up of 103 refrigerator-sized silver gray cabinets, occupies an area of nearly 1,000 sq m and weighs 155 tons, containing 6,144 Intel CPUs and 5,120 AMD GPUs, with a storage capacity of 1 PB.

"Tianhe can store four times of all the books housed in the National Library of China. If every one of China's 1.3 billion people takes a high-resolution digital photo, Tianhe can record all the photos," said Li Nan.

The giant machine, now housed in Changsha, will be installed in the National Supercomputing Center in north China's port city Tianjin at the end of this year.

It will be made available for domestic and overseas users in 2010.

China unveils its fastest supercomputer - dnaindia.com
 

amitkriit

Senior Member
Joined
Jul 17, 2009
Messages
2,463
Likes
1,927
Indian IT/Computer industry can keep surviving on borrowed "low cost" jobs, while our neighbour keeps creating new milestones. State of Indian R&D is absolutely pathetic when compared to China.
 

Yusuf

GUARDIAN
Super Mod
Joined
Mar 24, 2009
Messages
24,324
Likes
11,757
Country flag
I think people have forgotten about the PARAM. India made its supercomputer ages back. So people, dont keep criticizing our nation for nothing.
 

Daredevil

On Vacation!
Super Mod
Joined
Apr 5, 2009
Messages
11,615
Likes
5,772
Indian IT/Computer industry can keep surviving on borrowed "low cost" jobs, while our neighbour keeps creating new milestones. State of Indian R&D is absolutely pathetic when compared to China.
This is something that I don't like in some Indians who are disparaging towards their own country without any rhyme, reason or rationale. So, if we create a new super computer better than the Chinese one in next year and if there is no use for such fast super computer in our country then the purpose of making it is defeated. We don't need white elephants, which are expensive but of no use. We should build supercomputers if there is a need for it not just for the heck of it or for organ measuring contest.

FYI, making a supercomputer is not a difficult thing. You need housing space, more processing units, memory and you got it. In short, it needs money.
 

Vladimir79

Professional
Joined
Jul 1, 2009
Messages
1,404
Likes
82
Well Done.It's unbeleiveable that they can develope such kind of machine.
They didn't really make it, all they did was connect "6,144 Intel CPUs and 5,120 AMD GPUs" Any knowlegable person can hook up Xboxes to make a supercomputer. This is little different.
 

AkhandBharat

Regular Member
Joined
Aug 7, 2009
Messages
542
Likes
79
They didn't really make it, all they did was connect "6,144 Intel CPUs and 5,120 AMD GPUs" Any knowlegable person can hook up Xboxes to make a supercomputer. This is little different.
Vlad, It takes a little more to build supercomputers than to glue microprocessors together. There needs to be an OS which can execute tasks in parallel to utilize the full potential of the computing farm. Writing parallel algorithms is no joke. Take it from me. Also, I'd rather build a farm out of cell processors than xboxes :D

That being said, TCS (a private consulting company in India) had developed Ek (One) which was the world's 4th fastest computer in rankings 2 years ago when it was built. It had the computing power of 115 teraflops.

The chinese one claims 1 petaflop. They are not the first one to do it. The Japanese stole the show already :D Nonetheless, a nice supercomputer that will help China in genome sequencing, simulation and weather analysis etc.

CDAC is already building a petaflop supercomputer with an expected completion around 2010. There is no hurry.
 

Singh

Phat Cat
Super Mod
Joined
Feb 23, 2009
Messages
20,311
Likes
8,403
Country flag
Congratulations to China.

PS: If members feel their nation lacks achievements, go out and do something or shut up. No thread crapping please.
 

badguy2000

Respected Member
Senior Member
Joined
May 20, 2009
Messages
5,133
Likes
746
They didn't really make it, all they did was connect "6,144 Intel CPUs and 5,120 AMD GPUs" Any knowlegable person can hook up Xboxes to make a supercomputer. This is little different.
good phrases...it is always very enjoyable to read your comments. :)
 

tarunraju

Sanathan Pepe
Mod
Joined
Sep 18, 2009
Messages
9,080
Likes
40,077
Country flag
They didn't really make it, all they did was connect "6,144 Intel CPUs and 5,120 AMD GPUs" Any knowlegable person can hook up Xboxes to make a supercomputer. This is little different.
Vlad, that's pretty much how supercomputers are built these days, it's called the cluster topology. Using 'nodes'. Each node is a simple computer in itself. The nodes of Milkyway I (the name of this supercomputer), uses two Intel-made CPUs, and two ATI Radeon HD 4870 X2 graphics cards (four GPUs in all).

Here's the picture of Milkyway's node:


The only thing that dumbfounds us is the choice of Radeon HD 4870 X2. It sucks at general purpose GPU compared to monolithic GPUs from NVIDIA. Each Radeon HD 4870 GPU has an inflated shader compute power figure of 1.2 TFLOPs, which doesn't translate into application performance greater than GPUs from NVIDIA with half the rated SCP.
 

Vladimir79

Professional
Joined
Jul 1, 2009
Messages
1,404
Likes
82
good phrases...it is always very enjoyable to read your comments. :)
According to the official benchmark it is only running 560 teraflops which is a far cry from a petaflop. Lets wait for the next supercomputer rankings before we go swallowing CCP pills. Homemade Playstation SCs are already breaking 100 teraflops.
 

tarunraju

Sanathan Pepe
Mod
Joined
Sep 18, 2009
Messages
9,080
Likes
40,077
Country flag
Lets wait for the next supercomputer rankings before we go swallowing CCP pills. Homemade Playstation SCs are already breaking 100 teraflops.
When a shader core is crunching the most ideal problem (maybe 1 + 1 = ?), its throughput is measured (which is naturally high), and hence, a wonderful "TFLOPs" number is given out, which doesn't live up when faced with real-world problems. I do concede that the compute power of this supercomputer is inflated. For instance, each AMD GPU they claim to use (from that picture a Radeon HD 4870), comes with a rated compute power of 1.2 TFLOPs. They should add up to what, 6.14 PFLOPs right there, but you know the entire supercomputer lives up to less than 10% of that in the benchmarks, all those GPUs plus CPU cores put together.
 

Vladimir79

Professional
Joined
Jul 1, 2009
Messages
1,404
Likes
82
When a shader core is crunching the most ideal problem (maybe 1 + 1 = ?), its throughput is measured (which is naturally high), and hence, a wonderful "TFLOPs" number is given out, which doesn't live up when faced with real-world problems. I do concede that the compute power of this supercomputer is inflated. For instance, each AMD GPU they claim to use (from that picture a Radeon HD 4870), comes with a rated compute power of 1.2 TFLOPs. They should add up to what, 6.14 PFLOPs right there, but you know the entire supercomputer lives up to less than 10% of that in the benchmarks, all those GPUs plus CPU cores put together.
I am no programming guru so I would appreciate if you could explain to me how these supercomputers are tested for the global rankings.
 

masterofsea

Regular Member
Joined
Aug 6, 2009
Messages
243
Likes
1
This one is developed by the Computer Research Institute of CAS(Chinese Academe of Science).It mainly use Intel CPU and AMD GPU.
As a insider,I can tell you people that next year we will reveal a new model supercomputer.it have been developing by University of Science and Technology of China and ShuGuang Computer Corporation.It will use CPUs invented by China ourselves.
 

tarunraju

Sanathan Pepe
Mod
Joined
Sep 18, 2009
Messages
9,080
Likes
40,077
Country flag
I am no programming guru so I would appreciate if you could explain to me how these supercomputers are tested for the global rankings.
The list of the world's supercomputers is maintained by the Top500 project. It uses an industry standard benchmark suite called Linpack, and lists supercomputers based on their 'peak' and 'max' performance (in TFLOPs). The top 500 supercomputers are in the list. According to the June 2009 (latest) list, there are five supercomputers ahead of this Chinese one, assuming it does end up getting ~500 TFLOPs in Linpack. Since Linpack is an industry standard, it is safe to assume that the 'benchmark' which showed this supercomputer to have ~500 TFLOPs, is Linpack.
 

tarunraju

Sanathan Pepe
Mod
Joined
Sep 18, 2009
Messages
9,080
Likes
40,077
Country flag
This one is developed by the Computer Research Institute of CAS(Chinese Academe of Science).It mainly use Intel CPU and AMD GPU.
As a insider,I can tell you people that next year we will reveal a new model supercomputer.it have been developing by University of Science and Technology of China and ShuGuang Computer Corporation.It will use CPUs invented by China ourselves.
Perchance is it that Godsun/Loongsun (龙芯) CPU we keep hearing about? If so, it can be written off.
 

masterofsea

Regular Member
Joined
Aug 6, 2009
Messages
243
Likes
1
Vlad, that's pretty much how supercomputers are built these days, it's called the cluster topology. Using 'nodes'. Each node is a simple computer in itself. The nodes of Milkyway I (the name of this supercomputer), uses two Intel-made CPUs, and two ATI Radeon HD 4870 X2 graphics cards (four GPUs in all).

Here's the picture of Milkyway's node:


The only thing that dumbfounds us is the choice of Radeon HD 4870 X2. It sucks at general purpose GPU compared to monolithic GPUs from NVIDIA. Each Radeon HD 4870 GPU has an inflated shader compute power figure of 1.2 TFLOPs, which doesn't translate into application performance greater than GPUs from NVIDIA with half the rated SCP.
Do you know when the milkway-I was developed?
Please read the topic before you reply and do some search.
 

AkhandBharat

Regular Member
Joined
Aug 7, 2009
Messages
542
Likes
79
Its not as simple as linear addition of computing performance of nodes. The performance is measured by matrix manipulation in a period of time by the entire cluster, simply because 100% parallelism is like the mathematical infinity (unachievable). This is because bus speeds are the major hurdle in computing farms (cores operating faster than the bus speed).

Today's Symmetric multiprocessor based SIMD supercomputers' performance is measured by solving linear equations of the order n^3 (originally in 1979 an LU decomposition of matrix of size 100 of real double precision elements (not floating but double), but now can be 1000 or scalable). In Step 1 the entire matrix is decomposed into two matrices such that the product will eventually give the original non-sparce matrix of real elements. In Step 2, the decomposed matrices are used to solve the linear equations. The majority of operations are of the type of multiplying a scalar (single real number) to a vector (consider it to be an array of real numbers) and vector addition (adding an array of real numbers to another array).

More on this in the paper:
Linpack benchmark for supercomputing performance
 

masterofsea

Regular Member
Joined
Aug 6, 2009
Messages
243
Likes
1
According to the official benchmark it is only running 560 teraflops which is a far cry from a petaflop. Lets wait for the next supercomputer rankings before we go swallowing CCP pills. Homemade Playstation SCs are already breaking 100 teraflops.
Mr. Vladimir will build us a supercomputer with Xbox before the next supercomputer rankings before we go swallowing CCP pills.:goodstuff::goodstuff::goodstuff:
 

Latest Replies

Global Defence

New threads

Articles

Top