IBM Builds 10 PFlops Supercomputer
IBM will be building a 750,000 core supercomputer for the DOE's Argonne National Library.
Called Mira, the system is expected to deliver a maximum performance of about 10 PFlop/s which is about twice as fast as today's fastest supercomputer on the Top 500 list.
The new system will be built on IBM's BlueGene/Q technology and be capable of running about 10 quadrillion calculations per second, which is about 20 times the performance of Argonne's current Intrepid supercomputer. According to IBM, Mire will be able to do about the same number of calculations every second as every man, woman and child in the U.S. could do in one year if they were to do one calculation each second.
There was no information when Mira will be completed and there was no information how much this system will cost. However, IBM said that the supercomputer is a stepping stone on the path to an exascale supercomputer, which is expected to arrive in the 2020 time frame and will be at least 100 times faster than Mira. IBM said that such systems may be powered by "100s of millions of cores".
but can it stop me from destroying it with a sledge hammer?
no? then we humans are safe =]
365 days * 24 hours * 60mins * 60 seconds * 300,000,000 people =
9,460,800,000,000,000 calculations in one second
0.0
Wow. that would be double of what F@H doing right now (according to F@h stats, were running 5.2 native Petra flops or unless you want to count the x86 flops as 9.3 Petra Flops)
So if this thing were built tomorrow (which we know it wont) it could run the entire project with out our help. (in theroy)
Although by the time IBM finish building that thing, F@H will already be past the 10 pflop range.
Probably not!
The Bluegene/Q achieves roughly 1.7 GFLOPS/W. Going by this metric a 10 PFLOPS system would consume roughly 5.9 MW, which isn't all that much by supercomputing standards especially considering Mire's peak theoretical performance.
It's made up of 16-core quad-threaded PowerPC A2 based processors, each processor having 64 threads. So no, it isn't based on Cell, but even if it were, why would IBM get sued by Sony? The Cell architecture was developed and designed jointly by both companies, and IBM has already integrated it into many of its supercomputing designs.
correct me if im wrong, but aren't gpus ony really good at processing pre determined things faster?
i know there is one thing that makes a cpu a far better solution than a gpu, i just cant remember the thing.
jprahman, you do know that the 95W TDP is how much watts of heat a cpu and/or a gpu will release and not how much it consumes
Both the GTX 580 and the core i7's can consume more watts than there TDP's are. (not all the energy thats consumed by the hardware is lost as heat)
As for why there not doing a build of gpu's, you are right that gpu's are faster in flops but it's difficult to program for a gpu to run stuff and not all programs that need floating point calculations are that well suited for a gpu.
It also depends on energy usage and electical limitation as well. Yeah if a gpu can get stuff done faster, it could stop so it wouldn't use as much electricity..... Although how often do you here a super computer getting a break for completing a hard job? Rarely. So imagine all the cost of running that....
Now, there is one thing to re-look at of who making this supercomputer. It's IBM. So there going to be using something similar to the PS3 cell cpu they make or there servers powerXcell sever cpu that back in 2008 was already able to achive over 100 GFLOPS on eight SPEs.
Now whats different about the cpu's IBM has vs AMD/Intel? there cpu's are more like gpu's than cpu's but not fully gpu's. There like the hardware thats right in the middle of the 2. That at least according to the F@H programers that make and run the PS3 clients.
So what ever cpu be put in this, i going to have to guess that it will have higher than average cpu flops but lower than power usage than a GPU. This is if IBM makes a cpu with what the cell's have.