Japan's Petaflop Supercomputer

slashthedot writes "Japan has built the fastest supercomputer in the world. While the BlueGene/L contains 130,000 processors, Japan has managed to create the first Petaflop supercomputer, called MDGrape-3, with just 4808 chips, and it cost just $9 million to develop."

Comments

Progress

[Eightyford]

It now costs 15 dollars per gigaflop. In the early 90s, a million dollars per gigaflop was normal.

Incorrect chip count

[Bushcat]

The original article seems to be unreachable, so I can't read it, but the precis has the wrong chip count: It does have 4808 LSI chips, but it also has 19,122 Xeon processors.

Re:Incorrect chip count

[Savantissimo]

I read the article - don't waste your time. No doubt it's a cool machine, but the artile was the flimsiest puff-piece I've ever seen linked on Slashdot. Complete lack of technical detail, moron-level explainations of common terms - I feel stupider having read it.

Are there any good articles on this machine that anyone would care to share?

Re:Incorrect chip count

[rgravina]

This article here from Riken themselves has some more technical details:

http://mdgrape.gsc.riken.jp/modules/tinyd0/index.p hp

Say what?!?

[mosel-saar-ruwer]

Japan has managed to create the first Petaflop supercomputer, called MDGrape-3, with just 4808 chips...

FLOP = floating operation [per second].

PETA = 10 ^ 15, or "a quadrillion".

(10 ^ 15) / 4808 = about 207,986,688,852, which would indicate that each chip is running at several hundred TERA-hertz [and, even then, the machine would have to possess an operating system so efficient that it could consistently perform one floating point operation per clock increment, which seems extraordinarily unlikely].

Or is this an "analog" computer and are these "analog" FLOPS?

And no, I did not RTFA.

Re:Say what?!?

[hattig]

The Cell processor can do ~200 GFLOPS - not IEEE quality FLOPS however, however they're 'good enough single precision FLOPs' for it's target uses. This is probably why this new supercomputer won't get into the Top500 list, because it's very specialised and thus probably nowhere near as good at IEEE conformant calculations.

The Cell processor is not running at 200GHz. There's this concept called 'parallelisation', it's how your graphics card can do dozens, if not hundreds, of operations per clock cycle. In Cell's case it can do 8 (number of SPUs) * 4 (128-bit registers, SIMD) * 2 (units) = 64 SP FLOPS per clock cycle, and that's not including the PPU which has VMX128 and an FPU itself.

However make the Cell processor calculate IEEE conformant FLOPS, and it gets a double precision score of around 20GFLOPS. Still good though.

The above was from memory, details may vary, figures are roughly correct, YMMV, etc.

Re:Say what?!?

[Hollinger]

Yeah, it's a bit obvious that you didn't.

Quoting another link you can see how they reached these numbers (which I take issue with):

The following figure shows the block diagram of the MDGRAPE-3 chip. It consists of 20 force calculation pipelines, a j-particle memory unit, a cell-index controller, a master controller, and a force summation unit. The force calculation pipeline is the most important part of the chip which performs calculations of two-body forces such as Coulomb and van der Waals forces. Each pipeline performs 33 equivalent floating point operations per cycle when it calculates Coulomb force. Thus, when it operates at 250 MHz its performance will reach 165 Gflops with 20 pipelines. The chip also has the j-particle memory unit, which corresponds to the main memory of the CPU. Therefore, no extra memory is needed to attached with the chip.

- http://mdgrape.gsc.riken.jp/modules/tinyd0/index.p hp

With that answered, I'm confused. Another poster sent along that link which explains what Riken will do. I'm confused about that actually. Reading the page, based on the verb usage, either someone didn't understand future and past tense (possible, but unlikely), or they haven't built the entire box yet. Perhaps I'm reading a bit too much into it... it's quite possible that someone simply hasn't updated the website.

Based on the webpage, all of the calculations to reach 1 petaflop are based on theoretical peak performance measurements, extrapolated from the theoretical peak of a single special-purpose ASIC which has been built, but may or may not have been actually placed into a fully configured system. Nothing talks about measured benchmarks, and the OP's article contains the same theoretical extrapolated numbers.

Anyone know if they've actually built it?

~ Mike

For once the subtitle is right on

[John Muir]

ROFL at the "From the renders-a-million-tentacles-a-minute dept" ... nice choice!

Re:Our penis so small, your american penis so larg

[tomhudson]

"Show me the MFlops/Watt rating of this?"

No problemo!

The number of flops: (10 ^ 15) / 4808 = about 207,986,688,852 flops per chip, - from a previous poster.
The number of watts: 300,000 - from the manufacturers' site = 62 watts/chip
207,986,688,852 / 62 = 33,546,240 flops (33 MFlops) / watt.

Re:machines like this

[NewbieProgrammerMan]

If the resources are available to crack rc5, to do distributed based work on a cure for cancer, and crunch data captured from radio antennas in search of little green men from mars, then I think we have the know-how necessary get some thing like this up and running.

Well the examples that you mention are not really the same as "attempting to break software and search for problems long before release." If I understand these issues correctly: (1) (with apologies to crypto specialists) RC5 cracking required lots of CPU time to factor a big-ass number, (2) projects like Folding@Home aren't "looking for a cure for cancer," they're running (I think) quantum chemistry simulations to find out how certain molecules can act in certain situations, and (3) SETI@Home is looking for specific patterns in signal data. In all three of these cases, there's a few common (maybe not so simple) operations that need to be applied to a large set of data or initial conditions, and that's why they need lots of machines, or fast machines.

Figuring out how clever people will take advantage of a particular implementation of a web browser or TCP/IP stack is a completely different class of problem IMHO. Yeah, maybe there's some clever AI techniques that may simulate attack attempts, and maybe they could come up with attacks that nobody has thought of yet, but a really fast computer will not somehow magically solve these kinds of problems for us. There's a lot of hard science and software engineering that needs to be done first.

Specialised

[SamAdam3d]

The problem with that is that this computer is very specialised to molecular simulations. It can't very easily do other things, like seti or folding (okay, well, maybe that it can do). It was easy to design and cheap because it didn't have to be general purpose and adaptable, like BlueGene/L is.

Not Quite Progress Yet

[TubeSteak]

From Page 2 of TFA

No other supercomputer at the top of the rankings can muster so much calculating brawn on such a tiny budget. That's partly because MDGrape-3 relies on fewer chips and less circuitry than rivals. It's also because the chief scientist, It's also because the chief scientist, Dr. Makoto Taiji, working with only two other researchers, had plenty of help from Hitachi, Intel, and NEC subsidiary SGI Japan.

Those companies supplied the hardware -- Hitachi made the central processing unit, or CPU -- and absorbed part of the cost of building the machine. One measure of the MDGrape-3's ultra-efficient computing muscle is its cost per gigaflop (1 billion floating-point calculations per second), which Riken puts at $15.

Only if you're getting subsidised by 3 global corporations.

If it costs $15/gigaflop, then they would have paid... $15 million
A $6 million subsidy (40%) isn't small change.

Not comparable

[News for nerds]

Though the theoretical performance of this computer is higher than that of BlueGene and may have higher realworld performance too, you can't compare this supercomputer with BlueGene and other TOP500 supercomputers since it can't run LINPACK. It's just too specialized for its use.

Comparison MDGrape-3, BlueGene/L & Earth Simul

[yalla]

I compiled some quick facts which compare those three supercomputers and added pointers to other resources for your convenience:
http://www.bloglines.com/blog/ITnomad?id=126

Cheers, Alex.

Idiotic summery.

[imsabbel]

This computer, like all the previous (md)grape generations, is a central force potential calculation accelerator.

it does nothing but calculate 1/sqrt(dx^2+dy^2+dz^2)*variable, but really really often.

Grape 6, 5 years or so ago, was already running at 200Mhz, had a throughput of one force calculation per pipleline and 6 pipelines on once chip. So it counts as 1.2 billion force calculations, each being (1* inverse, 1 sqrt, 3 adds, 3 squares, 2 fmul, ect).
A lot of flops, but totally useless as general purpose computers.