backtop


Print 33 comment(s) - last by ksherman.. on Jun 23 at 11:49 PM

Second place on the list goes to a system with a max speed of 478.2 teraflops

IBM’s latest supercomputer – Roadrunner -- sits atop the Top500 supercomputers list that will be released at the International Supercomputing Conference in Dresden, Germany this week.

Its rise to the top of the list comes after it was able to break the petaflop barrier last week. News.com reports that Roadrunner was able to reach 1.026 petaflops, a bit over one quadrillion calculations per second.

The Roadrunner supercomputer dethroned IBM’s own BlueGene/L -- last year’s most powerful supercomputer -- pushing BlueGene/L to the number two spot on the list. BlueGene/L was able to reach 208.6 teraflops last year. This year it more than doubled its performance to 478.2 teraflops, but was still unable to match Roadrunner.

Roadrunner is based on the IBM QA22 blades that use an advanced version of the Cell processor found in the Sony PS3. The processing cores used in the Roadrunner are from AMD and make the machine the world’s first hybrid supercomputer.

Roadrunner is comprised of 278 refrigerator-size server racks and has 6,562 dual-core Opterons. IBM is the manufacturer of 210 of the 500 supercomputers on the Top500 list. Other well known makers with systems making the Top500 list include HP with 183 systems on the list and its top performer taking the number 8 spot on the list.

Sun has its Ranger system on the list at number 4, the Jaguar from Cray is number 5, Encanto from SGI is number 7, and Altix from SGI is number 10. Intel is the dominant processor in the supercomputer market powering 75% of all systems that made the Top500 list and 90% of ranked quad-core processor machines.

DailyTech covered the Roadrunner supercomputer last week.



Comments     Threshold


This article is over a month old, voting and posting comments is disabled

firestream?
By nosfe on 6/18/2008 3:03:40 PM , Rating: 3
quote:
In total, Roadrunner takes up 278 refrigerator-size server racks, and connects 6,562 dual-core AMD Opteron and 12,240 Cell chips.


and all that for 1 petaflop...
Now considering that a Firestream 9250 can do 0.2 teraflops(double precision) it would take 5k of these to match the roadrunner, that's only $5mil, of course you also need a ton of motherboards but i'm betting it would still end up cheaper than the $100mil it cost to assemble the roadrunner




RE: firestream?
By Amiga500 on 6/18/2008 3:16:27 PM , Rating: 4
Thats a bit of an apples and oranges comparison though.

This architecture is much more versatile than a series of GP-GPU cards.


RE: firestream?
By nosfe on 6/18/2008 3:28:02 PM , Rating: 2
is it? its just a bunch of cell processors, imo not that much different from GPGPUs, both are exotic compared to x86 also i'm talking about price/performance so its not apples and oranges


RE: firestream?
By spluurfg on 6/19/2008 10:04:10 AM , Rating: 2
quote:
is it? its just a bunch of cell processors, imo not that much different from GPGPUs, both are exotic compared to x86 also i'm talking about price/performance so its not apples and oranges


Yes, it's apples and oranges. If general purpose code were indeed faster on a graphics processor, you can bet that's what we'd be using right now. Also, consider the amount of memory bandwidth being fed to a Firestream GPU compared to an Opteron. It's an entire order of magnitude different. That would have very large implications when trying to run large numbers of these things in parallel and maintaining scaling performance.


RE: firestream?
By Hawkido on 6/18/2008 3:25:29 PM , Rating: 2
The Firestream is not capable of running in parallel on that scale... the Cell needs Opterons to handle all of the cross-traffic. While it is the Cell that is doing all the major leg work, the Opterons are just concerned with memory access and feeding the Cell processors, maybe some GPCPU functions as that is where the cell is weakest.


RE: firestream?
By nosfe on 6/18/2008 3:29:39 PM , Rating: 1
and whats stopping them from using opterons to handle the cross-traffic for the firestreams?


RE: firestream?
By drinkmorejava on 6/18/2008 11:12:04 PM , Rating: 2
Two things.

1) They don't have the hardware to integrate with hundreds of PCI channels each

2)The firestreams are like black boxes. Put some stuff get, get some different stuff out, ignore everything in between. It's great if you want to get one thing done independantly. But horrible when you try and scale it. I believe someone already mentioned this.

Also, a large part of the 100 million comes from have hundreds of engineers/support staff working on it since 2002. Pretty simple math actually.
for instance: 100 people*60k a year*6 years =36 mill


RE: firestream?
By emboss on 6/19/2008 2:43:24 AM , Rating: 2
Like I said in the FireStream article, it's all about interconnect. The latency between a FireStream and the main system is about 200 times that of the Cell (300 us vs 1.5 us). Or, in node-to-node terms, total latency between two cells in different nodes is about 5 us, compared to 1.5 to 2.0 us latency between two "traditional" nodes. If you used FireStreams or GPUs instead, total latency would be around 600 us node-to-node.

Given that the additional 1.5 us latency is already hurting some of the applications running on RoadRunner, the massive latency from the graphics card would simply wreck performance. Again, like I said in my other post, GPUs are great as long as you can fit everything into a single card.

In general, supercomputers are as much about interconnect as pure FP performance. Earth Simulator is a classic example: they basically built the biggest crossbar they could, and then stuck some computational capability on as an afterthought. There's some situations where it still can't be beat because of this design.


"Death Is Very Likely The Single Best Invention Of Life" -- Steve Jobs

Related Articles













botimage
Copyright 2014 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki