Fujitsu Launches 23 PFlops PRIMEHPC FX10 Supercomputer
Fujitsu has begun offering a new supercomputer, which the company claims can be scaled to a performance of up to 23.2 PFlops, which would make it the fastest system on the Top500 list and even exceed the upcoming BlueGene/Q-based Sequoia system.
To achieve this performance, the computer would be configured with a total of 1024 racks integrating a total of 98,304 computing nodes and 6 PB of memory. The base configuration has four racks with 384 processors.
What makes this particular announcement interesting is the fact that PRIMEHPC FX10 will be using the SPARC64 IXfx processor, the successor chip to the SPARC64 VIIIfx, which is currently used in the world's fastest supercomputer, K Computer system, in Japan. However, instead of eight cores, the SPARC64 IXfx integrates 16 cores, which means that Fujitsu's 23.2 PFlops supercomputer would boast a stunning 1,572,864 processing cores - more than twice the number of cores in the K Computer's 705,024.
According to Fujitsu, a single SPARC64 IXfx processor will deliver a floating point performance of 236.5 GFlops, about 85 percent more than the SPARC64 VIIIfx, which tops out at 128 GFlops. The company said that the computation efficiency is at about 2 GFlops per watt, which indicates that individual processors will run at about 115 watt at 1.85 GHz.
In comparison, IBM will use a 16-core PowerPC A2 processor in the 20 PFlops Sequoia system which will be installed at the Lawrence Livermore National Laboratory (LLNL) in 2012. IBM said that its A2 processor will consume only 30 watts; however, since the company states that the 98,304 computing nodes A2 system will reach just about 20 Flops, we would assume that the A2 is not quite as powerful as the SPARC64 IXfx.
- Useless Fact: Each Sperm Carries 37.5 MB of DNA Data
- A First Look At Google's Chrome Aura Interface
- The Best of Black Friday Deals 2011
- Rumor: Apple to Switch Back to Nvidia for 2012 GPUs
- IBM Patents Idea of Making Your Data Crappier Over Time
- Yikes! New Computer System Can Read Your Emotions
- LSI Announces SAS 12 Gb/s Interface
- Early Chrome Gets New Garbage Collector, Boosts Speed
- Razer Building Smaller Gaming Devices Besides Laptops
- IBM Files Patent for Putting Fire Extinguisher in a Computer
- The Deals for Cyber Monday 2011
- Pub Installs Pee-controlled Video Games in Bathrooms
- The First Intel Ivy Bridge CPU Clock Speeds and More
- AMD Intros Branded Desktop Memory Modules
- Is Intel Going to Kill its Celeron Processor Brand?
- The Deals for Cyber Monday 2011: Part 2
- Global HDD Supply 70 Million Units Short in Q4 2011
- MIT Researchers Invent Diode For Light in Optical Circuits
Guess not
But will it play cr.. No..
But not as power hungry. ~18% less compute power for ~3.8x less power consumption.
16 core cpu eh... interlaaaaagooooos.. did you miss the contract for this supercomputer, amd?
Tom's should do "best supercomputers for the money" for this kind of thing.
Japanese supercomputers tends to use japanase self made cpu's.
They could put this computer on fukushima nuclear power plant.
interlagos stands no chance against these CPU's
Flashback to 2007.
interlagos stands no chance against these CPU's
dum dum dum dum dum!
Just because its running now doesn't it will be in a week.
But will it play cr.. No..
Oh hell...it'd better be able to!!!
I love reading about all these new supercomputers. I also love reading articles showing which supercomputers (from the days of old) are being surpassed by my mobile phone (in time).
Oh hell...it'd better be able to!!!
Software rendered even...though most people that spout the crisis meme aren't old enough to know the days of having a software render as an option.
The trade-off between power and performance, is strongly in favor of IBMs PowerPC.
I'd rather buy 116,000 PowerPC A2s (consuming 3,480kW), which would equal the performance of the 98,304 SPARC64 IXfxs (consuming 11,600kW), and save tons of money with PowerPCs lower power consumption. Plus, it's more eco-friendly.
Saying they are offering a Scalable computer doesn't mean that it will come right away.
That and they would most likely not scale up to what they can scale it to.
Though a announcement of a Upgrade to a Super Computer should doesn't necessarily mean that the upgrade will go according to plan. Though most likely it could.
So keep your eyes open, because America's vying for the fastest super computer spot.
Good Luck Oak Ridge hope Linpack works well for you and your new Titan Super Computer!
Software rendered even...though most people that spout the crisis meme aren't old enough to know the days of having a software render as an option.
Ooh, that brings back memories of Asheron's Call and Homeworld in software mode on a friend's laptop. *cringe*
I'm thinking this thing could maybe even perform software based real-time ray tracing.
The article finishes too abruptly, I miss a discussion where apples are compared to apples: GFlops per watt to Gflops per watt, total cost against total cost, estimated total power consumption to estimated total power consumption. Pros and cons for each supercomputer. Blue Gene deliberately runs at low (half) frequencies because power consumption scales with the square of the frequency and performance scales linearly so they would rather lose some performance per core to save energy. The total cost of a supercomputer today has to factor in the energy consumption to run the thing and that is where Blue Gene usually wins.
Are you sure?
I think you missed something there.
Another one down the drain...... too much power.....
The trade-off between power and performance, is strongly in favor of IBMs PowerPC.I'd rather buy 116,000 PowerPC A2s (consuming 3,480kW), which would equal the performance of the 98,304 SPARC64 IXfxs (consuming 11,600kW), and save tons of money with PowerPCs lower power consumption. Plus, it's more eco-friendly.
Good point but I think the article is missing some key details that might change your mind.
1) The cost of both systems. Granted both would be extremely high in their most maxed out configuration but it the Fujitsu solution was say several hundred thousand less, might make power consumption a moot point.
2)Physical limitations of both systems. The article does state the number of racks for the Fujitsu system but not IBM (although I'm sure it exists elsewhere) and while the number of racks is a helpful number it does not state if that is a "turn key" setup, meaning, they don't say if other equipment is needed which could, depending on the equipment could be just as many racks.
3)And this may be the biggest point missed in the comparison, support. That's right folks even supercomputers come with warranties and tech support and at this level it is a HUGE factor. Now most of the places that have computers like this installed have employees that could, in theory, support them but when you have a multi-million dollar investment sitting in a data center. It is get really hard to justify to the top brass that they let a guy who they pay less than $60K a year replace a hard drive on such an expensive piece of equipment. Nope, they hire contractors, buy support contracts or call tech support and all of that comes from the company they bought the machine from, this is where IBM really makes their money.
I'm not trying to say your point is wrong, in fact I think I would go for the IBM if I was in the market for such a beast. Just because IBM has been doing the supercomputer thing for a while now and their support staff in my country is probably more plentiful than a Japanese based company.
Just my 2 cents
cool........that thing!
1) The cost of both systems. Granted both would be extremely high in their most maxed out configuration but it the Fujitsu solution was say several hundred thousand less, might make power consumption a moot point.
With professional-level computing, power consumption is NEVER a moot point. While it can safely be ignored for desktops, that's because they have this figure known as "idle TDP." With a server or supercomputer, that figure doesn't exist because these things don't run on idle, nor are they ever put on standby or shut down; they're always in use.
A 8,120 kW difference is staggering; even assuming a rather low cost of 10 cents per kWh of electricity, we're looking at $812US/hour extra for juice. That works out to well over half a million in just a single month, eating up the initial price costs and then some.
This is why peroformance-per-dollar doesn't mean anything for servers, and performance-per-watt does. To put it into perspective: each and every SINGLE watt that a server or supercomputer uses more, it'll cost an extra dollar per year of running. Then, of course, remember that datacenters are often bottlenecked by the amount of electricity that can be delivered by the grid in the first place: if your datacenter can only get 5 megawatts of available power, then that 11.6 MW SPARC-based system isn't an option, and you'd likely rather spring for the system that could get you the most performance that's possible in your constraints.
Had to post with a different nick ("anonymously"), could not post otherwise.
@seezur
You're right, I did the comparison based on the information provided in this article. I didn't dig any further, but I did take into consideration the following three Tom's articles I had read before:
http://www.tomshardware.com/news/i [...] 13400.html
http://www.tomshardware.com/news/k [...] 13904.html
http://www.tomshardware.com/news/s [...] 13979.html
These articles provide information about the rack configuration for IBMs PowerPC, and point out the enormous apetite of current supercomputers -especially the top ones- for power.
You can see why it was easy for me to pick out IBM as the better choise here. As for support, I believe IBMs support standards are on a very high level. I know nothing about support costs, though.
Your points are valid, but the difference in power consumption just blew everything else away. Energy efficiency is the deciding factor here in my opinion, and not just for the savings in the long run.
Also, @nottheking, great point made in the last paragraph.
The system looks nice too.