Top 500 supercomp list

[Janooo]

NV is used in #2 - Nebulae;
ATI is used in #7 - Tianhe-1;
Both built in China.

http://www.top500.org/lists/2010/06/press-release

 

[Idontcare]

What China doing with said computing resources? That's a lot of WoW accounts foo!

 

[Scali]

I think Intel's nightmare has just begun...

 

[Daedalus685]

I think Intel's nightmare has just begun...

Perhaps.. but Intel's share of the top 500 actually went up this time around.

 

[T2k]

What China doing with said computing resources? That's a lot of WoW accounts foo!

The Great Firewall...

 

[Genx87]

I think it is pretty interesting that Nebulea got pretty close to the #1 system with under half the cores and its peak is higher. Anybody have a break down of how many Tesla vs how many Xeons this thing runs?

Isnt Oak Ridge building a system based on Tesla as well? I suspect that should be pretty high up on the list.

 

[A_Dying_Wren]

Hmm... I really wonder what governments have that they aren't telling us though. Tin foil hats aside, it would be remiss for governments not to have some powerful computers possibly for decryption.

 

[Voo]

I think it is pretty interesting that Nebulea got pretty close to the #1 system with under half the cores and its peak is higher. Anybody have a break down of how many Tesla vs how many Xeons this thing runs?

Well the #1 has 224.162 opteron cores, while Nebulae used 120.640 cores, which is a rather useless metric without any further information.

But I've read a interesting article somewhere where they analyzed the data (after all we know how many cores the X5650 has, what it's GFLOPs performance is and we can assume some facts about the geometry of the whole infrastructure). I'll look if I can find it, but iirc it was around 4500 nodes with 2 processors and 1 gpu..
All in all extremely impressive, especially considering the fact that one SP is counted as one core.. though peak performance can't be really compared to other systems (much harder to get peak performance out of a GPU than a CPU)

@A_Dying_Wren: Even with such supercomputers it's not feasible to decrypt a message that's encrypted with a modern, strong algorithm - even if they would need only a few months for one message, doesn't sound like such a great idea, especially when there are much more interesting areas where you can use that performance (well maybe that's just me^^)

 

[WelshBloke]

This is a good link.

 

[Schadenfroh]

Good to see Oak Ridge still has the fastest supercomputer in the world and that my university still has the fastest supercomputer in Academia. Both are housed at ORNL. Likely more supercomputers there that we "do not know about."

 

[WelshBloke]

Good to see Oak Ridge still has the fastest supercomputer in the world and that my university still has the fastest supercomputer in Academia. Both are housed at ORNL. Likely more supercomputers there that we "do not know about."

Yeah, I'm thinking most governments arnt going to release much information about their stuff.

 

[faxon]

25TFlops and you can get on the list? i've seen distributed computing nuts with clusters that probably have that kind of throughput! to bad there systems arent all networked together as a single computational beast, would be interesting to see what the charts look like after that . i wonder where mark would score

 

[mv2devnull]

Yeah, I'm thinking most governments arnt going to release much information about their stuff.

Unless there is a requirement to publish what the tax-payers dimes are wasted on ...


I saw a comparison of the #1 and #2 somewhere. The "MW required" difference was impressive.

 

[nyker96]

I wonder if mine is on it :]

 

[faxon]

I wonder if mine is on it :]

currently mine can only hit 10.1TFlops theoretical peak in SP, but ATI GPUs dont have a linpack benchmark yet afaik so i cant test for sure :thumbsdown:. if i win the california super lotto you know what im gonna do with the winnings though :awe:

 

[Lonbjerg]

Good to see Oak Ridge still has the fastest supercomputer in a single benchmark of the number of supercomputers that have run this benchmark in the world and that my university still has the fastest supercomputer in Academia in a single benchmark of the number of supercomputers that have run this benchmark . Both are housed at ORNL. Likely more supercomputers there that we "do not know about."

Fixed...

 

[Idontcare]

25TFlops and you can get on the list? i've seen distributed computing nuts with clusters that probably have that kind of throughput! to bad there systems arent all networked together as a single computational beast, would be interesting to see what the charts look like after that . i wonder where mark would score

prolly not, cause once you start allowing aggregate flop metrics enter the list then everyone pimping a 50TFlops setup will find themselves still not on the list as the thousands of individuals out there with 100TFlop+ aggregate flop capabilities come out of the woodwork.

The rising tide would just make the minimum specs to get on the list all the higher, and still irrelevant since benching flops says nothing about the performance in the app of interest.

 

[f4phantom2500]

prolly not, cause once you start allowing aggregate flop metrics enter the list then everyone pimping a 50TFlops setup will find themselves still not on the list as the thousands of individuals out there with 100TFlop+ aggregate flop capabilities come out of the woodwork.

The rising tide would just make the minimum specs to get on the list all the higher, and still irrelevant since benching flops says nothing about the performance in the app of interest.

I can see it now:

"Mine's floppier!"
"No, MINE'S floppier!!"
"NO, MINE'S FLOPPIER!!!! CHECK OUT MY LINPACK!"


I was pretty stoked when I got my 4870, knowing that, in theory, my computer was faster than some on the top500 list from 2005/2006.

 

[HurleyBird]

Very interesting results. One thing that really stands out is how far the efficiency drops with the systems that use GPUs and/or other floating point accelerators.

The Jaguar system at oak ridge has attains 75.5% of its peak theoretical power.
Roadrunner, which uses Opterons and Cell processors actually reaches 75.7% efficiency.

The ATI and Nvidia systems take a bit of a plunge.

Nebulae, using Tesla accelerators, only has an efficiency of 42.6%
Surprisingly Tianhe-1, using commodity under-clocked 4870X2's, reaches 46.7% efficiency, actually beating Nebulae on the efficiency front.

There are a lot of unknowns here, but assuming that Nebulae uses 1-GPU/CPU like Tianhe-1 does, it's looking like ATI actually has a more efficient GPU-arch for Linpack, which is the opposite of what one would expect. Of course, programming and interconnects play a huge roll here as well, but you'd also expect that Nebulae would have advantages in those respects with this being a more modern system using supposedly easier to program for GPUs.

 

[jchu14]

Good to see Oak Ridge still has the fastest supercomputer in the world and that my university still has the fastest supercomputer in Academia. Both are housed at ORNL. Likely more supercomputers there that we "do not know about."

Hah, your UT overtook my UT. I use Texas Advanced Computing Center's Ranger for my research. I have fun telling people that I run my research code on one of the fastest computer in the world, but what they don't know is that my code can only use 16 cores at once.

Though, both of them will be overtook by UIUC's $200m Blue Waters with over 10 petaflop theroretical peak when it comes online in 2011.

 

[Idontcare]

I can see it now:

"Mine's floppier!"
"No, MINE'S floppier!!"
"NO, MINE'S FLOPPIER!!!! CHECK OUT MY LINPACK!"

I hear there's a blue pill you can give your computer to help with that

Hah, your UT overtook my UT. I use Texas Advanced Computing Center's Ranger for my research. I have fun telling people that I run my research code on one of the fastest computer in the world, but what they don't know is that my code can only use 16 cores at once.

Though, both of them will be overtook by UIUC's $200m Blue Waters with over 10 petaflop theroretical peak when it comes online in 2011.

When I was in need of serious compute power, a decade ago or so, I had a federal grant to use supercomputer time on one of the premier publicly acknowledged supers at the time but because of the way they sliced and diced the cpu allocations and priorities my specific application (which was highly multi-threaded and lent itself nicely to super-computer usage) actually completed it execution slower than if I just ran the program in linux on my lowly 4-node beowulf cluster at the time.

And the cost...my grant allocation was billed at something silly like $1000 per CPU hour and yet the performance on my $2k cluster which gave me 24/7 CPU time was higher and orders of magnitude cheaper.

Supers are great, very sexy for PR opportunities and the like, but when it comes to the actual customer (the end-user, which is NOT necessarily asked to make the purchasing decisions) who are attempting to run apps on them they get so loaded up and over-allocated that its hardly super-performance by any/every measure.

I bet the only time there is a single-app running on them is when they run linpack to get their top500 ranking and PR numbers verified. Otherwise its 1000 user trying to run in parallel.