NVIDIA Helps China to Supercomputer Crown

HardOCP News

[H] News
Joined
Dec 31, 1969
Messages
0
According to this article, China has taken the top spot on the supercomputer Top500 list with more than a little help from NVIDIA. The Chinese system beat out the U.S. based supercomputer from Tennessee for the honor of being number one.

Nvidia chips are powering a supercomputer that the graphics chip supplier claims has achieved the fastest speeds to date. The China-based "Tianhe-1A" has hit 2.507 petaflops, beating a system at Tennessee-based Oak Ridge National Laboratories.
 
That should be Nvidia and Intel according to the article I read.
 
... yea, but the crappy plastic will break after like two uses, and the screen is crap.
 
For Nvidia to claim that says the Chinese are running the machine in hybrid model. Wonder if theyr've programming the GPUs with PhyX? Would be interesting to see the software.

Top500 is down right now, how many CPUs does it have?
 
Good thing that CUDA thing isnt catching on....

This is a pretty big feat (although I wish it were a US system) thats a metric crap ton of computing power. hopefully china will use it for human advancement and not military.
 
Here is the nvidia press release http://pressroom.nvidia.com/easyir/customrel.do?easyirid=A0D622CE9F579F09&version=live&prid=678988&releasejsp=release_157

From the release - "The system uses 7,168 NVIDIA® Tesla™ M2050 GPUs and 14,336 CPUs; it would require more than 50,000 CPUs and twice as much floor space to deliver the same performance using CPUs alone."

Those unnamed CPUs are Intel Xeons.

Thanks for that info.

As you may know, GPUs typically operated much like vector processors so theoretically they could be 7 times faster than a scaler processor (using same clock speed). Thus the quote of 50K CPUs is reasonable considering you never reach the theoretical limit, clock speed, etc.

Applications using Linpac (the standard reference application) is normally a highly tuned one for the machine in question so you're not going to get near that number is real world apps. But that is the game they play. :D
 
For Nvidia to claim that says the Chinese are running the machine in hybrid model. Wonder if theyr've programming the GPUs with PhyX? Would be interesting to see the software.

Well they certainly aren't using PhysX. PhysX is a physics simulator for games. It lacks precision for "real" physics simulations and afaik isn't distributed (as in, can only use a single GPU).

Perhaps you meant CUDA, which is Nvidia's answer to GPGPU? They might also be using OpenCL.
 
Probably a nice 3.5-7M profit on that computer for Nvidia. Nice to see it getting it's super computing division up and running.
 
Well they certainly aren't using PhysX. PhysX is a physics simulator for games. It lacks precision for "real" physics simulations and afaik isn't distributed (as in, can only use a single GPU).

Perhaps you meant CUDA, which is Nvidia's answer to GPGPU? They might also be using OpenCL.

Thanks for the correction. I've not used the languages for the GPUs so I'm uneducated there. I just know there is typically a different language that is used for the GPU code. One builds portions of the calculation to run on the GPUs that best use them while other portions might be on the CPU.
 
I always thought nvidia was a bunch of commies with its logo anyway the way its meant to be played!!!!!!!!!! my game salways played better on ATI :D
 
Probably a nice 3.5-7M profit on that computer for Nvidia. Nice to see it getting it's super computing division up and running.

Typically nobody makes a profit on these machines. Most of the time vendors build them for bragging rights but I certainly don't know for sure. :D
 
...but can it play Crysis?

Now that I have that out of the way, I wonder how fast a machine like this could blow through folding work units, or how many WU's it could do in a day?
 
Probably a nice 3.5-7M profit on that computer for Nvidia. Nice to see it getting it's super computing division up and running.

I was wondering the financial effect for Nvidia.

The rough specs are here: http://www.nvidia.com/object/product_tesla_M2050_M2070_us.html
Up to 6GB of GDDR5 memory per GPU

Though I'm not sure if anyone knows if Nvidia supplied just the chips, or full boards with ram, pressumably a lot (1Gb+++) of ram since they're going to used for CPU stuff.


Anyone care to guesstimate Nvidia's revenue for this "one computer" ?

7,168 gpu's at... $___.00 each =
 
I wonder which tesla is set for physx, or if its CPU based.

But GPUs do make a load more sense for super computer based shenanighans, if you're having to do a mass load of simple tasks. Each tesla does have 448 "cores".

Still would be a load more badass if they watercooled them all with full cover blocks
 
I always thought nvidia was a bunch of commies with its logo anyway the way its meant to be played!!!!!!!!!! my game salways played better on ATI :D

<blink, blink>

95% of the games I play play exactly the same one my Radeons and my GeForces; a few have PhysX and a couple have show stopping bugs in one of the two drivers, more often ATI.

So WTF mate? Or are you just a ridiculous fan boy?

If you want to get into company politics and ethics, well thats another discussion... But play wise, if things are working as they should, you really ought not notice a difference.
 
I was wondering the financial effect for Nvidia.

The rough specs are here: http://www.nvidia.com/object/product_tesla_M2050_M2070_us.html


Though I'm not sure if anyone knows if Nvidia supplied just the chips, or full boards with ram, pressumably a lot (1Gb+++) of ram since they're going to used for CPU stuff.


Anyone care to guesstimate Nvidia's revenue for this "one computer" ?

7,168 gpu's at... $___.00 each =
Well it's going to be priced under what the 50,000 CPU equivelent would be. Assuming those are 500$ CPUs, then the most they could have sold the GPUs is (50,000-14336)*500/7168 = 2,487$ per. So taking it down to 1,500 per and 300$ cost is still 1200K. Considering it's exactly the same as a 480 except they've got a little more Ram, the 300$ cost is probably too high, probably closer to 200$.

I wouldn't be surprised if they got them nearly free of charge for the pure marketing value.
See the math above, they can make the super computer for much cheaper than a CPU equivelent and still make money hand over fist.
 
The cost itself is irrellevant, I mean Cisco for example loves give away hardware for free just so they can use you as a reference customer for future sales. Hence why I wouldn't be surprised if Nvidia gave this kit away just so tay can say, "hey we are powering the most powerful cluster on the planet, wanna buy a few Tesla cards?"

In my example they made over 7 million dollars, and still were able to say they did it for 1/2 the cost of using CPUs. I don't see the point of shorting your self the 7 million when you can say that.
 
How long until GPUs are on export ban list to specific countries?

Not much help since they're built there, huh?
 
Applications using Linpac (the standard reference application) is normally a highly tuned one for the machine in question so you're not going to get near that number is real world apps. But that is the game they play. :D

I think the big problem is that linpack is SOOOOO optimized for CPU that it doesnt give the GPU enough credit.

There are a lot of codes out there that will run x times faster on a few GPUs than on a lot of Top500 clusters

Typically nobody makes a profit on these machines. Most of the time vendors build them for bragging rights but I certainly don't know for sure. :D

I am willing to bet otherwise, they may not have made "normal" money but I am sure they still took a profit

I was wondering the financial effect for Nvidia.

The rough specs are here: http://www.nvidia.com/object/product_tesla_M2050_M2070_us.html


Though I'm not sure if anyone knows if Nvidia supplied just the chips, or full boards with ram, pressumably a lot (1Gb+++) of ram since they're going to used for CPU stuff.


Anyone care to guesstimate Nvidia's revenue for this "one computer" ?

7,168 gpu's at... $___.00 each =

If they are Tesla C2070s they MSRP ~$3500/pc

Well it's going to be priced under what the 50,000 CPU equivelent would be. Assuming those are 500$ CPUs, then the most they could have sold the GPUs is (50,000-14336)*500/7168 = 2,487$ per. So taking it down to 1,500 per and 300$ cost is still 1200K. Considering it's exactly the same as a 480 except they've got a little more Ram, the 300$ cost is probably too high, probably closer to 200$.


See the math above, they can make the super computer for much cheaper than a CPU equivelent and still make money hand over fist.

Dont forget about the power savings

Lets assume that they payed residential cost @ $.10/kwh that equates to tad of $14mill/year to run? right? (((4000kwh X $.10)x24hrs)x365 days)=14016000

They say that CPU cluster would have been 3x the cost to run. Thats $42mill/yr

Obviously they dont pay res rates but its still relative and we haven't even got into cooling costs (let the fermi heat jokes roll in...)

Where? TSMC isn't in China.
Taiwan
 
Um...

Taiwan is Chinese. It's just not part of the PRC... Sorta. The People's Republic of China (PRC) still claims the territory governed by the Republic of China (ROC/Taiwan) and to be the true government of China. Of course the ROC thinks that it is sovereign and the true government of China.

Good times, and quite likely the reason that the PRC might nuke Los Angeles some day if they decide to reintegrate Taiwan by force and the US follows up on it's defense agreements with the ROC.

But this is all beside the point, the question should be where was the Foxconn factory that assembled these? :D
 
So... You're saying that the Republic of China isn't China? :D

But that still doesn't answer which Foxconn facility actually assembled the boards. Was it Shenzen, or Foxconn City in Lenghua, or elsewhere?

The headquarters are in the ROC not the PRC so it could be either mainland China or Taiwan.
 
This machine actually used to be red. The Tianhe-I was originally powered by 2,560 dual-GPU ATI Radeon HD 4870 X2 GPUs.

They upgraded to Tesla and renamed it the Tianhe-IA.
 
Can anyone give me a laymans example of how they use these super computers to model weather or model military applications like nukes or things of that nature. I've tried to google this info myself but there is not a much info out there that can be read easily.
 
That's it!

No more Green team for me! Helping the Chinese and dissing the U.S.

ATI only from now on!
 
That's it!

No more Green team for me! Helping the Chinese and dissing the U.S.

ATI only from now on!

I support this line of thinking, except that virtually ANYTHING you buy in the "technology" world has China's hands in it at some point.
 
Back
Top