Japan Wants to Build 10 Petaflop Supercomputer

deepexplorer writes "Japan wants to gain the fastest supercomputer spot back. Japan wants to develop a supercomputer that can operate at 10 petaflops, or 10 quadrillion calculations per second, which is 73 times faster than the Blue Gene. Current fastest supercomputer is the partially finished Blue Gene is capable of 136.8 teraflops and the target when finished is 360 teraflops."

Comments

BlueGene/C will be finished soon

[Raul654]

BlueGene/L is the fastest super computer at the moment; however, BlueGene/C (which, for the record, I'm working on as part of my PhD) will be finished very soon (it was supposed to be out of the foundry by the end of August, but the project is running slightly behind schedule). I'm told there are, as yet, no plans to publish any performance benchmarks.

Re:BlueGene/C will be finished soon

[hazzey]

Is there a reason why they aren't even close to alphabetical order?

Do the letters stand for something else?

Re:BlueGene/C will be finished soon

[mfloy]

I'm obviously not as close to the source as you are, but aren't there other computers in the BlueGene family being worked on also, that will be faster than both L and C? The quote of the Japanese computer being 73 times faster than BlueGene/L, but by the time the computer is actually built I will be suprised if it is faster than the current incarnation of BlueGene. IBM has been doing some great work as far as supercomputing goes, working on it must be a blast.

Re:BlueGene/C will be finished soon

[unit00]

Roman numerals perhaps? C=100 L=50

Re:BlueGene/C will be finished soon

[Raul654]

Linux (w/ a custom made compiler that my group has already written)

Re:Mommy, mommy

[Anonymous Coward]

Are you taking grudge from WWII ? Get over with it. They still kick ass in technology and make better AS WELL AS cheaper cars.

If you want faster , here ya go.

[zymano]

Superconducting supercomputer. Too expensive but maybe need to build one to see how they work.
http://www.hq.nasa.gov/hpcc/insights/vol6/supercom .htm

Using 'general' processors is cheap but the wrong direction according to the best supercomputer expert from Stanford. He designed some cray computers.

http://content.techweb.com/wire/26802955

Re:teaflops

[nwbvt]

I had enough trouble getting past the grammar issues with that sentence:

"Current fastest supercomputer is the partially finished Blue Gene is capable of 136.8 teaflops and the target when finished is 360 teraflops."

Seriously, what do the editors do here? They don't check the writing, they don't check the accuracy of stories, and forget about it if you want them to post a correction to something...

Re:It's all about the $'s

[Jeff DeMaagd]

The other current top performers are PowerPC, Itanium and Opterons. Maybe not vanilla processors but not custom either.

BlueGene's PPC chips ARE custom for that line of computer, though VTech's Mac cluster is pretty much off-the-shelf.

Itanium isn't custom they are not hard to get, just that there isn't much demand. I think they are kind of nifty, though not competitive for general server use, might be OK for supercomputers, and has high-availability features not found in Xeon and Opteron.

I'm not sure if there is anything special about Opterons now, other than having more hypertransport links, and being special binned parts to take higher temperatures and consume less wattage than a comparable Athlon64, much like Xeon is to Pentium4. I think 1xx Opterions are basically the same as Athlon64.

Its a government makework project

[patio11]

I'm sure you guys have heard of our propensity for building bridges here? Including long bridges to islands with no real need for them, built in multiples sufficient to carry the entire population of the island off of it at a single time? Which are then built to withstand typhoons and earthquakes (well, OK, THATS not irrational). This is the same thing, except for the tech industry. And the US government does the same thing -- NASA and a good deal of the Department of Defense R&D fund are basically slushfunds to keep engineers employed in the hope that they come up with something useful in the meantime (and I would be remiss if I didn't point out that pork is well-appreciated come election time).

I don't really know why we love gigantic computers, though. I live in a prefecture which is Japan's answer to rural Iowa and we built a 1,300 node distributed supercomputer without any idea of a feasible application to run on it -- we ended up computing a few zillion solutions to N-Queens before mothballing the project (I was hoping for enough CPU time to take the world record back from the real supercomputer at the Japanese university that currently holds it, but unfortunately it was not to be).

Re:Its a government makework project

[demachina]

Much of the U.S. fixation goes back to the signing of the Comprehensive Nuclear Test Ban Treaty in 1996. The U.S. had a bunch of powerful labs full of top scientists whose job in life was to build and test nuclear weapons. This treaty pretty much put them out of business. Clinton distracted them by giving them millions of dollars to build gigantic supercomputers. The goal was to simulate nuclear explosions, predict how the U.S. nuclear stockpile would age and insure it would still work if the need arose without ever testing it ever again. They use to prove this by taking one out and setting it off in Nevada to make sure it still worked. Now they write simulations. Maybe the are very good at those simulations and they can in fact insure the nuclear arsenal is sage and potent. Unfortunately if they never set one off again they will never now if their simulations are any good. They might just be wasting billions of dollars.

In many respects the national labs are like NASA, they are high tech job programs for deep thinkers who would be dangerous if they were unemployed like their counterparts in Russia.

So they build giant computers, and hopefully figure out useful code to run on them though its not clear if they do have anything useful to run on them. There are always weather sims and protein foldings to do.

The worst problem is the tyranny of Moore's law. They take years to complete and by the time they are fully operational they are obsolete so you just start building a new one.

You wonder how people designed engineered marvels like the first fission and fusion bombs, Apollo and the SR-71 back in the day when they had next to no computing power. Now we have this extraordinary computing power but we have real problems building interesting things in the real world. The Shuttle made massive use of CFD, CAE etc but its a complete lemon. We keeping doing massive simulations of nuclear bombs but we never actually set any off and really don't even want them anymore. Well thats not true the Bush administration is in fact trying to restart development of new nukes and in fact want to build one for busting bunkers and caves. If they manage to get it built not only will the test ban treaty be out the window but the U.S. will start using them as a matter of routine in conventional wars and maybe just to take out a suspected nest of terrorists here and there. Maybe all this computing power will help make them in to exceptionally good tactical weapons which will get a lot of mileage.

Re:To put 10 Petaflops in perspective

[iendedi]

I just finished reading this article. There is a wonderful excerpt from that article that I would like to quote:

"A petaflop is roughly a human brain-second. Peta is equal to a million gigaflops or a million gigahertz Pentium processors. So we're crossing to a transition of computing power (equivalent to what's) in your head. What will we do with it, or it with us?

I am left with the sense that we should be abandoning flop-talk and simply move to a new measurement: Human brain-second is really alluring. Let's just shorten it to brainsec.

So this new Japanese supercomputer is running at a whopping 10 brainsecs!!! Imagine, you could simulate about 9 people or 47 slashdotters in that supercomputer (some of the power would be required to manage the simulatioins).

Seriously though, AI research will go mainstream with the first supercomputer that can process at greater than 1 brainsec.

Definition of supercomputer getting stretched

[heroine]

Nowadays the supercomputer contest is just a matter of who can buy the most Opteron PC's and Cisco routers from Newegg and connect them. You might as well buy a few million DVD's from Best Buy and say you have the world's largest hard drive.

Eventually small countries will connect all the computers of their entire population with distributed clients and call that the world's largest supercomputer.

This business of entering a command, waiting a minute for zillions of nodes across a slow network to start, and waiting another minute for all the nodes to finish is hardly what supercomputing used to be.

It would be more interesting to see who does the most work with the least latency or who does the most work with the simplest programming model. Anyone can write a massively parallel program to utilize every Opteron in the world but a computer which can do the same work sequentially seems like a much bigger step forward.