100 Million-Core Supercomputers Coming By 2018

CWmike writes "As amazing as today's supercomputing systems are, they remain primitive and current designs soak up too much power, space and money. And as big as they are today, supercomputers aren't big enough — a key topic for some of the estimated 11,000 people now gathering in Portland, Ore. for the 22nd annual supercomputing conference, SC09, will be the next performance goal: an exascale system. Today, supercomputers are well short of an exascale. The world's fastest system at Oak Ridge National Laboratory, according to the just released Top500 list, is a Cray XT5 system, which has 224,256 processing cores from six-core Opteron chips made by Advanced Micro Devices Inc. (AMD). The Jaguar is capable of a peak performance of 2.3 petaflops. But Jaguar's record is just a blip, a fleeting benchmark. The US Department of Energy has already begun holding workshops on building a system that's 1,000 times more powerful — an exascale system, said Buddy Bland, project director at the Oak Ridge Leadership Computing Facility that includes Jaguar. The exascale systems will be needed for high-resolution climate models, bio energy products and smart grid development as well as fusion energy design. The latter project is now under way in France: the International Thermonuclear Experimental Reactor, which the US is co-developing. They're expected to arrive in 2018 — in line with Moore's Law — which helps to explain the roughly 10-year development period. But the problems involved in reaching exaflop scale go well beyond Moore's Law."

Comments

100 Million?

[Itninja]

Can't we just start calling this a 'supercore' or something? When the numbers get that high it kind of goes beyond what most people can visualize. Like describing how hot the Sun is....let's just says it's "exactly 1 Sun hot".

Re:100 Million?

[MozeeToby]

How about 1 million cores being a mega-core. So the proposed supercomputer would be a 100 mega-core computer.

Re:100 Million?

[Jeremy Erwin]

The core? The surface? The corona?

Re:100 Million?

[Anonymous Coward]

Because even though the number's "effect" on you diminishes as it goes up, doesn't mean it is still significant. There's a reason Engineers use quantitative instead of qualitative.
How do you tell the difference between hot and really hot or really really hot?

Really.

How about the difference between 10, 20 and 30?

10

Which gives you more information?

Re:100 Million?

[Yvan256]

Let's just make sure it's 1 000 000 cores and not 1 048 576 cores... let's not make that mistake again.

Re:100 Million?

[Yvan256]

The definition of "supercomputer" changes as time goes by. Today's cellphones are yesterday's supercomputers.

Re:100 Million?

[_KiTA_]

Let's just make sure it's 1 048 576 cores and not 1 000 000 cores... let's not make that mistake again.

Re:100 Million?

[Yvan256]

Just because CS has been abusing a system for over four decades doesn't make it right.

Re:100 Million?

[Stargoat]

I suspect that we will end up calling this a heuristically designed processor. Or something similar....

Re:100 Million?

[Stupid McStupidson]

Fuck everything, we're going 200 million cores

Re:100 Million?

[Anonymous Coward]

Loggers don't measure in yards, they measure in board-feet.

Why can't computer science use a measurement that's based on powers of 2, which is exactly what they work in?

Re:100 Million?

[Anonymous Coward]

The definition of "supercomputer" changes as time goes by. Today's cellphones are yesterday's supercomputers.

Weird. My cell phone today is the same as my cell phone yesterday. And don't even get me talking about my so-called "supercomputer" tomorrow.

Re:100 Million?

[Archangel Michael]

One Million Cores and one Sun hot mentioned in the same post, coincidence? I think not!

Re:100 Million?

[Anonymous Coward]

Can you translate that in "Library of Congress's"?

Re:100 Million?

[oldspewey]

When loggers measure in board feet, they call them board feet not yards. When drive makers (or whoever else) measure in kibibytes, they call them kilobytes not kibibytes.

Re:100 Million?

[Yvan256]

It's not about using a power of 10 vs power of 2, it's about using the SI units for the larger units. A kilo means 1000, not 1024.

We're at a point where we have hard drive manufacturers getting sued by users who are confused by all this damn mess.

Here's some information about the subject.

Re:100 Million?

[aztracker1]

Far more computer science types wind up working with money (base-10) than anything base-2 or base 16.

Re:100 Million?

[Anonymous Coward]

Loggers don't measure in yards, they measure in board-feet.

Oh, that's where log4j is going wrong then ... it doesn't even log distance :-(

Re:100 Million?

[Anonymous Coward]

+1, Agreed.

Having multiple conflicting definitions of a term renders that term meaningless.

Re:100 Million?

[Mikkeles]

'When loggers measure in board feet, they call them board feet not yards....'

Board-feet measure volume; yards measure length.

Re:100 Million?

[TheKidWho]

Well, what was missing from the article summary is that this computer is going to be built using nVidia GPUs, not CPUs for the majority of computing...

Although really, with the way Fermi is shaping up, it is turning into a very specialized CPU.

Re:100 Million?

[Thud457]

Can you translate that in "Library of Congress's"?

yes. It generates the same amount of heat as burning 37 Libraries of Congress.

Re:100 Million?

[HPLovecrack]

As long as computers work on a base-2 system then the numbers wont align to a base-10 system. Using base-2 numbers for magnitudes or system numbers makes sense. Why put 3 cores in a CPU ( Looking at AMD) if it takes the same amount of binary notation to point to 4. This argument stands for your 1 000 000 cores.

Re:100 Million?

[Gilmoure]

I prefer Caesar's work with LoA temps.

Re:100 Million?

[JesseMcDonald]

You use SI prefixes with SI units. The 'byte' is not an SI unit; it's not even the most basic representation, being a group of eight bits. If you insist on using base-ten units in combination with bytes then you're essentially arguing for layering a base-ten system on top of a base-two one.

So far as I know there is no designated SI unit for information. Following the pattern of the other SI units, however, the best choice would be the bit. If you want base-ten measurements, then, you should use "kilobit", "megabit", etc., which unambiguously use the SI prefixes, official unit or not. The non-SI term "megabyte" will never unambiguously mean "10^6 bytes", and trying to make it so just renders the term useless for any purpose requiring precision.

Re:100 Million?

[wirelessbuzzers]

The Indians already have a word for 10 million cores.

Re:100 Million?

[sexconker]

CS abused nothing.
KB means 1024 bytes, and it always will.

KB is not K.
KB is not stepping on the toes of any SI units.
SI units are not sacred.
SI units are not enforceable by law.
SI units step on their own toes and are ambiguous themselves.

Anytime you see a b or a B after a K, M, etc. scalar multiplier, you are talking about bits or bytes and are using 1024 instead of 1000. It is not confusing. It is not ambiguous.

It's the fault of storage device marketers and idiot "engineers" who didn't check their work, made a mistake on some project, and refuse to admit it that the "confusion" exists.

Furthermore, classical SI scalars are used for measuring - bits are discrete finite quanta - we COUNT them. Would you like a centibyte? TOO FUCKING BAD.

The scalar of 1000 was chosen out of pure convenience. The scalar 1024 was chosen out of convenience, and was made a power of 2 because of the inherent nature of storage with respect to permutations (how many bits do I need to contain this space at this resolution? how much resolution and space can I get from this many bits?) and because of physical aspects relating to the manufacturing and design of the actual circuits.

CS has a fucking REASON to use 1024.
SI does not have a fucking reason to use 1000.

There is more validity in claiming that all SI units should be switched to 1024 than there is in suggesting KB mean 1024 bytes.

"But everything written before the change will be ambiguous!!!" yet you SI proponents tried to shove that ibi shit into CS (and failed miserably, thank you) despite the fact that it would cause the same fucking problem ("Does he mean KB or KiB?" "When was it published?" "Uh, Copyright 1999-2009" "Uh...").

In short, 1024 is correct, 1000 is wrong.

Re:100 Million?

[Anonymous Coward]

Like describing how hot the Sun is....let's just says it's "exactly 1 Sun hot".

Dude 1: "She's soooo hot!"
Dude 2: "Yeah, Man, she's 1 Sun hot"
Dude 1: "No way, she's at least 2 Suns"

Re:100 Million?

[sexconker]

Fail.
I'll eat my hat when drive marketers measure storage in anything but the made up notion of a scalar factor of 1000.

When drive marketers measure in 1000s of bytes, they call them kilobytes. They created half of the confusion. The other half came from non-CS engineers who made a mistake on a CS-related project because they were ignorant and refuse to admit their ignorance.

Re:100 Million?

[GargamelSpaceman]

With this many cores, don't they cease being cores, and become more of a smudge?

Re:100 Million?

[sexconker]

I'm still waiting for the hedonistically designed processor.

Re:100 Million?

[sexconker]

Although really, with the way Fermi is shaping up, it is turning into a very specialized paper weight.

Fixed that for you.

Re:100 Million?

[Quantumstate]

SI has a very good reason. Try converting from mm to km. It is trivial. Now try your number *1024*1024, this clearly isn't as friendly.

Therefore the reason we use 1000 in SI is because we work in base 10 and it makes the numbers easy.

Re:100 Million?

[Anonymous Coward]

Yes, numbers in computers go from 1, 2, 4, ..., 1024, 2048, 4096, ..., 1048576, ..., etc. Nobody is arguing against that.

Nobody is debating wether or not computers work in base 2 for some areas (such as RAM, addressing, etc), so stop bringing that up, it's not a valid argument for making a kilo being equal to 1024. Not everything in a computer is base 2, such as the ethernet port being 10/100/1000.

The problem is exactly that "1024 was chosen out of convenience". A kilo means 1000 and that's all there is to it, you cannot change facts.

As for "causing the same fucking problem", we're already there. Fixing it now would mean that at least from this point forward we'd know for sure what people meant, that's why the new units were proposed. If I say "1 kibibyte" you know it's 1024. If I say "1 kilobyte" you can't be sure I mean 1000 or 1024.

CS always have such stupid problems. If I give you the date "10/11/12", what the hell does it mean? First of all there's the Y2K problem with old data, second you have the MM/DD/YY vs DD/MM/YY vs YY/MM/DD problem (and ISO 8601 fixes this problem beautifully, YYYY-MM-DD).

Just because there is old, set ways to do something doesn't mean it's the right way.

Re:100 Million?

[Anonymous Coward]

Tough to tell - are you arguing for or against the adoption of a new system of prefixes?

Re:100 Million?

[mdwh2]

it's about using the SI units for the larger units

A bit isn't an SI unit.

And if you reply, answer me this: how much RAM does your computer have? Are you going to answer 2GB, or 2.147483648GB?

Re:100 Million?

[Hatta]

If I say "1 kibibyte" you know it's 1024.

I then also know that you're a twit.

If I say "1 kilobyte" you can't be sure I mean 1000 or 1024.

I know you mean 1024, unless you're a scummy hard disk manufacturer. The only reason to ever use kilobyte to mean 1000 bytes is to mislead people.

Re:100 Million?

[mdwh2]

...and just to add, if you're so obsessed with applying SI terms to information units, why do we have bytes? 8 bits to a byte? That's hardly in line with SI is it, surely you should replace all references to "byte" with "0.8 decibits"?

Re:100 Million?

[ari_j]

Therefore, cell phones on January 2, 2018 will have 100 million processor cores.

Re:100 Million?

[cashX3r0]

its gonna take a great product like vista to take advantage of this multicore technology.

Re:100 Million?

[Yvan256]

It's 2 Gibibyte, of course (1073741824 bytes).

Re:100 Million?

[Yvan256]

There's no reason to refer to SI units for a byte, since "byte" means 8 bits. There's no such thing as "bytemeters" or "bytegrams".

I find it funny that people confuse the whole "base 2 vs base 10" with "SI units". A byte is 8 bits, a kilobyte is 1000 bytes and a kibibyte is 1024 bytes. It's simple.

Re:100 Million?

[HappyEngineer]

The term megabyte will always be ambiguous. Half the people are confederates who yell "don't mess with history!" (meaning computing history) and the other half are union people who yell "don't mess with standards!" (meaning SI unit standards).

In my opinion you can all megafuck off. The <prefix>-byte terms became ambiguous once people started arguing over the definitions. Even a gigawhine of arguments will not make them one kilowhit less ambiguous. A meter is only a meter because there is near universal agreement. Without agreement the term would be ambiguous.

The terms are ruined. Feel free to cry a terabucket of tears if you wish.

Re:100 Million?

[b4upoo]

But can it play chess?

Re:100 Million?

[Anonymous Coward]

Let's just make sure it's 1 048 576 cores and not 1 000 000 cores... let's not make that mistake again.

I have no problem with you using 1 048 576, as long as you use the binary prefix (mebi, in case you don't know). If you insist on using mega, then you display your ignorance, and perpetuate the mistake.

Re:100 Million?

[Surt]

So SI is about being lazy with math? How about we just stick with the lazy 1024 bytes to the kilobyte then.

Re:100 Million?

[Anonymous Coward]

Nuts, anybody have the conversions formula here?

Re:100 Million?

[Nefarious Wheel]

CS abused nothing.

KB means 1024 bytes, and it always will.

KB is not K.

What a beautifully engineered rant. Correct, too.

But a bit of philosophy here - SI units were designed to measure physical characteristics, and to do so in a way that made computation easy. A litre is a kilogram of water, standard composition, and conversions between items measured are fairly simple as a result.

Bits are mathematical constructs emulated in a physical symbol, a switch state. They are mathematical abstractions, not measurements. And as PP has implied, the concept of bringing them together - "centibytes" - is as silly as asking the number of library of congresses there are in a bicycle. Easy computation requires one Kb to be 1024 bits, not 1000. Hats off to you!

(or perhaps, given your name, you can leave your hat on...)

Re:100 Million?

[Nefarious Wheel]

I would suspect Heuristically programmed ALgorithmic processor. When you get 9000 cores together, anyway.

I'm sorry, Dave...

Re:100 Million?

[sexconker]

That reason is convenience.
1024 has convenience as well as actual reasons.

Did you not read my post?

Re:100 Million?

[Stargoat]

I'm glad someone got the joke.

Re:100 Million?

[Yvan256]

Feel free to cry a terabucket of tears if you wish.

What do you mean? An african or european terabucket?

Re:100 Million?

[Rising Ape]

You use SI prefixes with SI units. The 'byte' is not an SI unit; it's not even the most basic representation, being a group of eight bits. If you insist on using base-ten units in combination with bytes then you're essentially arguing for layering a base-ten system on top of a base-two one.

Kilo and Mega are SI terms. The point of the prefix is it means the same thing no matter what it applies to. WIthout that, there's no point in having standardised prefixes. Electron-volt isn't an SI unit either, does that mean that keV and MeV should mean something other than 1000 and 1 million eV respectively? Similarly barn, parsec etc. The confusion would be endless. Computer technologists should have made up their own damn prefix if they wanted something that meant 1024. And I don't know what your point is anyway - there's nothing "basic" about any of the SI base units, they're all arbitrary to some degree, and represent quantities not neatly represented by base-10 anything. Doesn't stop them being usable as SI units.

Your point about using "bit" as the fundamental unit of information makes a lot of sense. Would make it much easier to work out in your head how long it would take a 2 Gb file to transfer over a 100 Mbps connection, for example.

Re:100 Million?

[Anonymous Coward]

No, SI is about counting on your fingers, genius.

How many fingers do you have? Do you have 10? Do you naturally think in a base-10 (that is, {0, 1, 2, 3, 4, 5, 6, 7, 8, 9})? Do you know your tens as well as you know your native language?

1000 = 10 * 10 * 10. Easy. Friendly.

And I'm sure it's much easier to remember 1000 meters to a km rather than 1760 yd to a mile or 5280 feet to a mile.

inb4someonewithlessthan10fingerscallsmeaninsensitiveclod

Re:100 Million?

[bandmassa]

How about Peak Oil coming by 2015? That's going to put the brakes on the number of cores in our computers. Won't matter what we call it. When the energy runs out, there'll be no food, let alone new computers.

Re:100 Million?

[lennier]

"Now try your number *1024*1024, this clearly isn't as friendly."

1*2^10 * 1*2^10 = 1*2^20 . Easy peasy.

Oh, you meant in base 10? What's that?

Re:100 Million?

[jimmydevice]

Nobody in the States uses SI units. That's the problem.

Re:100 Million?

[gregben]

Knowing what I know, and when I knowed it,
I'd say AMD has 4 cores on that die and one
just blowed it.

Re:100 Million?

[sFurbo]

Anytime you see a b or a B after a K, M, etc. scalar multiplier, you are talking about bits or bytes and are using 1024 instead of 1000. It is not confusing. It is not ambiguous.

So, whenever you have a unit prefixed by M, it means 1000000, except in these two particular case. How is that not confusing and/or ambiguous?

CS has a fucking REASON to use 1024. SI does not have a fucking reason to use 1000.

CS have a reason for 1024, but none for wanting it to be called k. SI have a reason for wanting every k to mean 1000. It is fine if you want prefixes which fits the use of one particular field, but don't just take something welldefined and give it a conflicting meaning. That is just asking for trouble, which is exactly what you got.

Re:100 Million?

[Anonymous Coward]

Sorry, but we already lost this battle by proxy (you and I had no vote). Take it up with the IEEE and NIST if you have a problem with the new units. Personally I've already made my peace and learned to accept the new standard.

IEEE 1541 recommends KiB (kibibytes) for 2^10 and MiB (mebibytes) for 2^20.

NIST guidelines (Warning: PDF) require use of IEC prefixes KiB, MiB ... (and not kB, MB) for binary byte multiples.

Re:100 Million?

[TheRaven64]

You're arguing against yourself. Using powers of two is more computationally convenient because, unlike most physical quantities, we always deal with bytes in power of two terms. Why? Because we index bytes by binary values. We use base 10 for SI units not because it makes multiplication and division easy, but because it makes logarithms easy (which, in turn, makes multiplication and division easy).

Take a look at that shiny new 64-bit computer. How much memory can it address? 2^64 bytes. Base 2 logarithm is 64. Now we have a value that's convenient to work with in base 10, so we can go kilo (54), mega (44), giga (34), tera (24), peta (14), exa (4) and get the answer; 2^4 exabytes, or 16 exabytes. Trivial to work out without a calculator; you can even do it in your head without needing a piece of paper. Now try doing the same with base-10 units.

This isn't limited to memory. Filesystems index by blocks, for example. Each block is 512 bytes. How much data can you store in a 32-bit filesystem? 2^32 blocks, each of which is 2^9 bytes, means 2^41 bytes, or 2GB. Again, easy to do in your head with base 2 logs, now try it with base 10 logs.

Re:100 Million?

[rubi]

No need to, the core count isn't being done in powers of 2, modern humans use a base-10 counting system (maybe the fingers have someting to do with it, but I don't know).

Re:100 Million?

[rolfwind]

A "byte" wasn't defined by th SI community at the time, "kilo", "mega", etcetera were. The CS community took over and twisted generally accepted terms for their own little niche. Not every scientific branch should be free to do that, just to fuck with their peers in all other fields.

Re:100 Million?

[mdwh2]

Sure, that was the term invented by some people to replace a pre-existing term, but I'm curious if you really say "Gibibyte" to people?

Re:100 Million?

[mdwh2]

But which is the base unit? If they're really SI units, there can only be one.

So either a bit is the base unit, in which case you should refer to a byte as 0.8 decabits (I meant deca, not deci), or a byte is the base unit, and you should refer to a bit as 0.125 bytes.

A byte is 8 bits, a kilobyte is 1000 bytes and a kibibyte is 1024 bytes. It's simple.

8 bits to a byte, and 1000 bytes to kilobyte? And then kibibytes on top of that? Simple? That's nothing like the SI units that I know.

Re:100 Million?

[mdwh2]

These prefixes were around before the SI people used them - they come from Greek words. And when pre-existing words are used, it's commonplace that more than one field may use them, with different meanings. Furthermore, it's not true that they redefined terms - kilo etc only have their meaning when used as a prefix for an SI unit. Since the bit or byte aren't SI units, there is no redefinition.

But later on, some people did decide to redefine kilobyte etc to fudge and conflate it with the SI meaning.

Not every scientific branch should be free to do that, just to fuck with their peers in all other fields.

The only people who "fuck with their peers" are the ones who decided to redefine the terms kilobyte, etc, just to cause confusion.

Re:100 Million?

[JesseMcDonald]

Kilo and Mega are SI terms. The point of the prefix is it means the same thing no matter what it applies to.

So a megaplex is exactly 1000 "plexes", then? (Whatever a "plex" is...) And a microprocessor is exactly one millionth of a processor? And "nanotechnology" refers to something exactly one-billionth the scale of a "technology"? Face it, these prefixes both predate SI and have a long history of being used in ways other than that defined for SI units.

Electron-volt isn't an SI unit either, does that mean that keV and MeV should mean something other than 1000 and 1 million eV respectively?

They could, but precedence with regard to this specific unit argues for the use of SI prefixes. Precedence in the case of "bytes" argues otherwise.

And I don't know what your point is anyway - there's nothing "basic" about any of the SI base units, they're all arbitrary to some degree, and represent quantities not neatly represented by base-10 anything.

My point is simply this: would you rather discuss information as integer multiples of one bit, which has strong precedence in information theory, or standardize on "0.125 bytes" instead due to the what amounts to a historical accident? It only makes sense to talk about bytes when the bits are already organized into power-of-two groups, in which case it also makes sense to use powers of two for larger quantities (which are likewise organized into powers of two, for various reasons).

Computer technologists should have made up their own ... prefix if they wanted something that meant 1024.

They did. Unfortunately they're particularly bad at choosing names, and came up with prefixes no one wants to use. ("Mebibyte"? Seriously?)

Come up with a better set of names and people might actually switch. In that event, of course, the historic terms ("megabyte" and the like) will probably have to be abandoned as too ambiguous for practical use, so it wouldn't really be a "win" for either side.

Re:100 Million?

[sexconker]

M means many things in the SI world.

It means 1 million.

It stands for mass, with capitalization distinguishing between the minor body and the major body in a simple system, or subscripting on a capital M being used to denote various bodies.

It stands for meters, with no regard ever given to capitalization.

K?

1000.
Spring constant.
Kinetic energy.
Fucking Kelvin.

SI units are trash at being unambiguous.
Stylization and capitalization and subscripting/superscripting bullshit does not translate at all to plain text or anything written by hand.

CS didn't use K or M. They used Kb, KB, Mb, and MB. There is no confusion.

Re:100 Million?

[sFurbo]

So, there is ambiguity in the SI system for variables and constants, the prefixes and units have less ambiguity and are left behind, we need to make some there?
If anybody other than a CS uses M as a prefix for a unit, they mean that it is a factor 1000000 larger. If anybody other than a CS use k as a prefix for a unit, it they mean 1000. If a CS uses them in that manner, he means 2^20 and 2^10. How is this unambiguous?

I understand the need for prefixes for the powers of 2, but why use the same letter as something that is almost, but not quite, the same?

Re:100 Million?

[sexconker]

"So, there is ambiguity in the SI system for variables and constants"
Yes.

"the prefixes and units have less ambiguity"
No, they're all part of the same ambiguity.

"we need to make some there?"
That was the thinking behind adding in the "ibi" bullshit.

It's unambiguous because a computer scientist does not use K or M or G or anything in isolation. It's always attached (directly or in context) to a b or a B. KB is not K, and thus there is no ambiguity.

Re:100 Million?

[Anonymous Coward]

And computers work in base 2 which makes the numbers easy.

Re:100 Million?

[ioshhdflwuegfh]

M means many things in the SI world.

Not really. In SI, M stands for mega, which is 1 million.

It stands for mass, with capitalization distinguishing between the minor body and the major body in a simple system, or subscripting on a capital M being used to denote various bodies.

Not true. SI only regulates expression of units, not how symbols are used in mathematical formulas. Also, at least in physics and astronomy, M and m in formulas usually do denote masses, but no "major"/"minor" masses, whatever that might mean, but larger (M) and smaller (m) mass. This notation is not SI nor is it formalized.

K?

In SI system, K stands for Kelvin, unit of temperature. K does not stand for kilo (which is 1000), ok? SI symbol for kilo is k.

Stylization and capitalization and subscripting/superscripting bullshit does not translate at all to plain text or anything written by hand.

This is bullshit alright.

Re:100 Million?

[treeves]

No, I'm going to answer 4GB.

Re:100 Million?

[Anonymous Coward]

LOL.
So it's k, M, g (since G is the gravitational constant), ... ?

M for meter and M for mega and M for milli.
There is NO WAY you can disambiguate that.

Re:100 Million?

[noinoii]

i use netbook @noiinoii http://adsbookmarks.com/

Who's President, Future-boy?

[pete-classic]

As amazing as today's supercomputing systems are, they remain primitive

Wait, what? You lost me. Are you from the future? How can you describe the state of the art as "primitive"?

-Peter

Re:Who's President, Future-boy?

[NoYob]

As amazing as today's supercomputing systems are, they remain primitive

Wait, what? You lost me. Are you from the future? How can you describe the state of the art as "primitive"?

-Peter

Oh, that's silly! He's psychic, of course. He can SEE into the future!

Re:Who's President, Future-boy?

[Yvan256]

Forget the president, ask for the winning lottery numbers for the next 20 years!

Re:Who's President, Future-boy?

[mcgrew]

My cell phone is a supercomputer. At least, it would have been if I'd had it in 1972. Rather then being from the future, he, like me, is from the past and living in this science fiction future when all that fantasy stuff like doors that open by themselves, rockets to space, phones that need no wires and fit in your pocket, computers on your desk, ovens that bake a potato in three minutes without the oven getting hot, flat screen TVs that aren't round at the corners, eye implants that cure nearsightedness, farsightedness, astigmatism and cataracts all at once, etc.

Back when I was young it didn't seem primitive at all. Looking back, GEES. When you went to the hospital they knocked you out with automotive starting fluid and left scars eight inches wide. These days they say "you're going to sleep now" and you blink and find yourself in the recovery room, feeling no pain or nausea with a tiny scar.

We are indeed living in primitive times. Back in the 1870s a man quit the Patent office on the grounds that everything useful had already been invented. If you're young enough you're going to see things that you couldn't imagine, or at least couldn't believe possible.

Sickness, pain, and death. And Star Trek.

Re:Who's President, Future-boy?

[Z00L00K]

You can still predict that some tech is primitive.

When a computer develops a mind of it's own in a logical manner it's starting to reach the human level and we can start to discuss if it's primitive or not. If it starts to reproduce on it's own it's time to be careful.

Re:Who's President, Future-boy?

[David Greene]

Wait, what? You lost me. Are you from the future? How can you describe the state of the art as "primitive"?

Pretty easily, actually. There are lots of problems to solve, not the least of which is programming model. We're still basically using MPI to drive these machines. That will not cut it on a 100-million core machine where each socket has on the order of 100 cores. MPI can very easily be described as "primitive," as well as "clunky," "tedious" and "a pain in the ***."

How do we checkpoint a million-core program? How do we debug a million-core program? We are in the infancy of computing.

Re:Who's President, Future-boy?

[Anonymous Coward]

-Peter

Peter! Hey how's it going! We have been looking for you to help out this poor unfortunate guy Paul over here. Ok truth is.. there are millions of Pauls I hope you don't mind spreading your wealth around.

Re:Who's President, Future-boy?

[vtcodger]

***How do we debug a million-core program?***

What is this "debugging" thing you speak of? If you are asking how we will test software for a million core system, we'll do it the same way we always have. We'll get a trivial test case to run once, then we'll ship.

Re:Who's President, Future-boy?

[aztracker1]

No nausea? WTF, I've gone through two surgeries where they put me out in the past 5 years, I was nauseous after both of them.

Re:Who's President, Future-boy?

[aztracker1]

Sarah Conner?

Re:Who's President, Future-boy?

[pete-classic]

Wow, a bunch of people didn't get what I thought was a simple point.

I understand that the current state of supercomputing will seem primitive at some point in the future. In fact, my post is predicated on that notion.

But words mean things. At some point in the future everything about our current state of culture and technology will seem primitive. Describing the current state-of-the-art as primitive is meaningless. That approach can be applied to any topic equally.

Let me illustrate by counter-example. "The practice of medicine in parts of sub-Saharan Africa remains primitive." See how I'm creating a contrast that conveys meaning? Such a contrast only exists in the summary if the author has some frame of reference extending into the future.

Also, I was making a Back to the Future reference.

*shrug*

-Peter

Re:Who's President, Future-boy?

[Jekler]

I think of our supercomputing systems as primitive in an analogous way as cavemen wouldn't end up with a rocket thruster if they just throw enough logs on a fire.

Without more advanced software designs and some type of revolutionary system architecture, more cores ends up only being slightly better than linear progression. They're primitive in that our supercomputers are seldom more than the sum of their parts.

Re:Who's President, Future-boy?

[Anonymous Coward]

Disclaimer: I know you're making a joke but I'll pretend it wasn't so I can start a flame war.

You seem to imply that testing removes the necessity of debugging.

Just because a program worked once doesn't mean it will always work. And TDD reduces problems, it doesn't make them disappear.
 

Re:Who's President, Future-boy?

[turgid]

When a computer develops a mind of it's own in a logical manner it's starting to reach the human level and we can start to discuss if it's primitive or not. If it starts to reproduce on it's own it's time to be careful.

That's not directly related to computing power per se. A computer 100 000 000 times as powerful as today's, running today's software will still not have developed a mind of its own. It'll just be very, very fast indeed.

Re:Who's President, Future-boy?

[mcgrew]

I guess I was lucky, or you were unlucky. Or my anesthesiologist was better than yours. I've had a hemorrhoidectomy, cataract surgery, and a vitrectomey in the last ten years and it was all painless and nausea-free. Well, the hemorroid surgery was painful the next day, and the vitrectomy was hell on my spinal arthritis, but that's only because I couldn't raise my head except ten minutes an hour for two weeks. Fifty years ago the detached retina would have completely blinded an eye, and cataract surgery wasn't possible until 1949.

Whan my friend Charlie woke up in the recovery room she didn't show any signs of nausea either.

Re:Who's President, Future-boy?

[aztracker1]

I'm pretty sensitive in general to most medications... so it's probably me as much as anything.

Re:Who's President, Future-boy?

[mcgrew]

Don't feel bad, my daughter is allergic to almost everything.

Re:Who's President, Future-boy?

[sexconker]

Whan my friend Charlie woke up in the recovery room she didn't show any signs of nausea either.

Did "she" show any signs of having a new, lumpy dick made from arm muscle?

Re:Who's President, Future-boy?

[ebh]

I always remember to ask the anesthesiologist for anti-nausea meds before going into the OR.

Re:Who's President, Future-boy?

[ebh]

Instead of zero-day exploits, it'll have zero-millisecond exploits.

Re:Who's President, Future-boy?

[mikael]

The simplest way is to be able to split the algorithm into pipeline stages then split each pipeline stage up into blocks of tasks that can be subdivided to cores as necessary. Then you only need to debug/test a single task to get a multi-threaded pipeline stage implemented.

Re:Who's President, Future-boy?

[Anonymatt]

Dude, I liked reading your journal entry.

Re:Who's President, Future-boy?

[mikael]

Some primitive medicine was more advanced that medicine 100 years ago.

The process Ancient Egyptians used to make beer, also happened to create antibiotics (tetracycline) at the same time.

For helping relieve head injuries, trepanning was done at the far back of a dry cave. There wasn't any organic matter for bacteria to feed on, the cave walls and roof would be free from dust, and a freshly chiseled flint blade would also be free of bacteria.

Re:Who's President, Future-boy?

[Surt]

You can only bet the next lottery number for $1, and you might have to share the prize. Knowing the next president you can place bets with larger payoffs.

Re:Who's President, Future-boy?

[Anonymous Coward]

You should have had a vasectomy, rather than a vitrectomy, then maybe you wouldn't have a child.

Re:Who's President, Future-boy?

[caywen]

"The future is here" is probably the most false, yet constantly true statement ever.

Re:Who's President, Future-boy?

[Xyrus]

Especially considering the software developed for such platforms. Most models run on supercomputers are written in Fortran. If there is an real software engineering, it's more of an afterthought. The core of some of these models still contain F77 code, with F77 styles and program structure complete with detailed variable names such as iqxn4. They're not well documented. They're very fragile. Porting to a new system (even if running just a different flavor of Linux) can take weeks to months because of the way implementation is done. Sometimes just going to a new version of the same compiler will cause things to break. And sometimes you need to be an advanced linux just to get everything to run, let alone compiling and linking.

Models up to this point have been ad hoc conglomerations of pieces of code individuals have written and managed to stick together. These often are individuals who were not trained to be software engineers and know enough about programming to be dangerous. Sometimes these codes were never intended to be used outside of a research paper (but were used anyway). They models work, but not much else.

In fact, if you want a prime examples of how software should NOT be written, there's a number of scientific models that fit the bill (complete with 10,000 line subroutines).

As far as supercomputing goes, we are just barely emerging from the "highschool hacker" days. If you want an application that scales to these levels effectively, you need to have some people who know what they're doing get involved. You also need the tools, which have been sorely lacking in this area.

Ah well. Bitch and moan.

~X~

Re:Who's President, Future-boy?

[jd]

MPI is indeed clunky. Most parallel libraries out there are. (MPI and PVM are built on the idea that every node has a shell and that you can run programs by rsh-ing onto that node and running the program.)

There are more elegant ways of running things. For example, rsh and a remote shell implies that the node has a fairly good-sized OS running, which means you're eating up system resources before you do anything.

MOSIX seems a better starting point. If you can shunt already-active processes over a network, you don't need a shell, you don't need userspace stuff that's going to be stagnant most of the time. (The active process needn't be a full program, it can be something that pulls in a dynamically-loaded object that has the real program in it when it gets to the destination. In fact, it's better if it ISN'T the program, since then you don't run into the issue of whether kernel threads should migrate or not.)

For collective operations, you really want to use multicasting (as then you don't waste bandwidth or create excessive latency). Scalable Reliable Multicasting of some sort - NACK-Oriented Reliable Multicasting being the popular form - would be a good starting point. The destination nodes can calculate any node-specific details faster than a master node can, since the master node would have to do so sequentially using no information not present in the collective message.

Message passing itself is so-so as an approach. You can see from the popularity of the original MACH-based HURD that there are definite limitations to the approach. In the end, whether you're talking about remote I/O, remote procedure calls or other remote operation, you're ultimately talking about sending and receiving data, where that data can be delivered to an active thread, trigger the start of a new thread, change the state of a mutex/futex/semaphore or perform some other very basic task.

Now, different approaches to parallelism work on different assumptions. MPI assumes you've a well-defined master node and well-defined slaves in almost the reverse of the traditional client-server model. Pi-Occam assumes you've well-defined channels, where channels can be fixed or mobile. I =think= channels are all point-to-point, as Occam was originally developed with the Transputer's mesh topology in mind and still carries some of the assumptions.

There are "metaschedulers", which seek to schedule operations over a whole cluster as though it was a single virtual machine. It's a good concept, there is no real distinction between N physical machines being one logical server, or one physical server being N logical machines.

However, "metaschedulers" tend to be run on a master node. If we're running on a MOSIX-type cluster, there isn't any need for a master node. Indeed, there isn't a vast need for a single metascheduler. Rather, you'd want many mini metaschedulers, each concerned with their local area, which interacted with the local scheduler and each other.

Re:Who's President, Future-boy?

[mcgrew]

If I'd had a vasectomy you wouldn't be around, boy. BTW your mother is lousy in bed, I only fucked her once. Now get off my lawn, dickweed.

Re:Who's President, Future-boy?

[mcgrew]

Actually her name is Sharlene, but when she was young she coulnd't pronounce it correctly and it came out as "Charlie", and it stuck. Or so she says.

But no, she doesn't have a dick.

Re:Who's President, Future-boy?

[mcgrew]

Thank you! Glad you liked it, there are more. I have to warn you, though, that some of them are NSFW and at least one of them made at least one slashdotter cringe.

Those are older journals from another (now unused) ID; I'd forgotten the password to his one, and an email to help@slashdot.org with the email address I used with this older one got it back for me.

yeah but yeah but

[Anonymous Coward]

my old trusty VIC20 ftw

Sorry - I can't help myself

[RPGonAS400]

Can You Imagine a Beowulf Cluster of These?

Re:Sorry - I can't help myself

[sherpajohn]

Only if they run Linux and can render Natalie Portman covered in hot grits faster than my imagination already does....woohoo!

Re:Sorry - I can't help myself

[Anonymous Coward]

Can You Imagine a Beowulf Cluster of These?

/me wanks furiously

Re:Sorry - I can't help myself

[Narpak]

Only if they run Linux and can render Natalie Portman covered in hot grits faster than my imagination already does....woohoo!

Your imagination must be quite primitive. Mine did that instantly upon reading your post.

Re:Sorry - I can't help myself

[dubbreak]

Only if they run Linux and can render Natalie Portman covered in hot grits faster than my imagination already does....woohoo!

Your imagination must be quite primitive. Mine did that instantly upon reading your post.

Oh yeah. Well my imagination automatically made her naked and petrified in addition to being covered in grits.

Re:Sorry - I can't help myself

[Anonymous Coward]

Your imagination must be quite primitive. Mine did that instantly upon reading your post.

Heck, I rendered it after reading "Natalie". Boy, I miss those memes. In Soviet Russia, old korean people render Natalie Porman covered in hot grits for you!.

Re:Sorry - I can't help myself

[Surt]

You only think it was instantly due to your low resolution windows timer.

Limits on simulation.

[140Mandak262Jamuna]

The programming techniques and mathematical formulations needed to take advantage of such very large number of processors continue to be the main stumbling blocks. Some kind of simulations parallelize naturally. Time accurate fulid flow simulation for example is very easy to parallelize and technically you can devote a processor for each element and do time marching nicely. But not all physics problems are amenable to parallelization. Further even in the nice cases like fluid flow, if one tries to do solution adaptive meshing, no uniform grids etc, the time step slows down so much the simulation takes too long even on a 100 million processor machine.

The CFL condition that limits the maximum time step one can take shows no sign of relenting. Score has been Courant (the C in CFL) 1, Moore 0 for the last three decades.

Re:Limits on simulation.

[radtea]

The CFL condition that limits the maximum time step one can take shows no sign of relenting. Score has been Courant (the C in CFL) 1, Moore 0 for the last three decades.

Yeah, I always get a laugh out of people who think that we're ever going to beat down turbulent flow with higher resolution. It's vortices all the way down, and no matter how clever your implicit scheme you still have to be able to propogate information through the grid at less than the speed of sound to prevent numerical shock waves from blowing up your solution. Regularization schemes that throw away information are good, but that reduces the value of going to higher resolutions.

So I'm doubtful that we'll be predicting the weather, or the climate, with significantly greater accuracy ten years from now than we are today. Some problems just don't yield to brute force very well, although one would hope that at least the higher resolution models will conserve energy and have free boundary conditions in place of today's frequently artificially fixed ones.

Heisenberg is reputed to have said at the end of his life, "I have two questions I want to ask God: 'Why relativity?', and 'Why turbulence?' I'm really hoping He'll be able to give me an answer on relativity..."

Re:Limits on simulation.

[David Greene]

Further even in the nice cases like fluid flow, if one tries to do solution adaptive meshing, no uniform grids etc, the time step slows down so much the simulation takes too long even on a 100 million processor machine.

That's true in general. However, techniques like dynamic scheduling can help. Work stealing algorithms and other tricks will probably become part of the general programming model as we move forward. More and more of this has to be pushed to compilers, runtimes and libraries.

Re:Limits on simulation.

[Anonymous Coward]

No, fluid flow parallelizes horribly - you have to calculate the time steps in order.

Re:Limits on simulation.

[140Mandak262Jamuna]

What? Explicit global time stepping scheme parallelizes badly? My understanding is that it parallelizes nicely but the solution time explodes because if you halve the grid spacing, the mesh becomes eight times bigger and time step becomes half and your solution takes 16 times longer. Solution scaling at O(N^4) kills you. The only thing worse than this is probably Cramer's Rule to solve simultaneous equations. (Recalling from the classic text book in CFD by my guru Dale Anderson with Tannehill and Pletcher.)

Re:Limits on simulation.

[jstults]

The programming techniques and mathematical formulations needed to take advantage of such very large number of processors continue to be the main stumbling blocks.

True.

The CFL condition that limits the maximum time step one can take shows no sign of relenting. Score has been Courant (the C in CFL) 1, Moore 0 for the last three decades.

There are more suitable methods for stiff, multi-scale problems (implicit time integration, preconditioning, multigrid) that remove those CFL constraints and alleviate the convergence problems (ill-conditioning) with large, high-resolution grids. They may be harder to parallelize, but they make those big problems more tractable. I think most spectral/pseudo-spectral global circulation models (the summary mentions climate modeling) use some sort of implicit time-stepping at least.

Re:Limits on simulation.

[jstults]

Yeah, I always get a laugh out of people who think that we're ever going to beat down turbulent flow with higher resolution. It's vortices all the way down,

But at some point you ought to be able to model the 'sub-grid' eddies (homogeneous, isotropic), so you just need a calibrated turbulence model for those.

and no matter how clever your implicit scheme you still have to be able to propogate information through the grid at less than the speed of sound to prevent numerical shock waves from blowing up your solution.

This sounds like word-salad to me, if the scheme is a stable discretization of the governing equations there's no need to worry about propogating info at less than the speed of sound. That's why implicit schemes allow you to take large time-steps, maybe you're thinking of CFL limits on explicit schemes based on cell-size?

Regularization schemes that throw away information are good, but that reduces the value of going to higher resolutions.

Higher orders and higher resolutions still pay-off even with filtering and artificial viscosity.

So I'm doubtful that we'll be predicting the weather, or the climate, with significantly greater accuracy ten years from now than we are today.

Me too. The limit on our predictive ability is probably the resolution and accuracy of our observations of the system rather than the models, the models are already at significantly higher resolution than the measurements, which should give us cause to re-evaluate where we put our money: more simulation or more satellites?

Re:Limits on simulation.

[speed of lightx2]

Not true. That is only true on explicit discretizations, CFL doesn't apply for implicit problems. On the other hand, you do have to invert a large matrix, but there are tools for doing that with large sparse matrices.

Re:Limits on simulation.

[jstults]

(Recalling from the classic text book in CFD by my guru Dale Anderson with Tannehill and Pletcher.)

TAP is a great comprehensive CFD reference (I still refer to it often), but it is a bit dated, so its coverage of some of the more modern stuff is limited or missing, I'd highly recommend Blazek's book as a supplement. Not as comprehensive, but good coverage of the modern developments.

Re:Limits on simulation.

[jstults]

...you have to calculate the time steps in order.

Foiled again by that damned 2nd Law (shakes fist in air)!

Re:Limits on simulation.

[The Master Control P]

It's only vortices and turbulence until you get to some characteristic length dictated by viscosity. Then again, good luck getting to that point with any computer in the near future.

Re:Limits on simulation.

[140Mandak262Jamuna]

In implicit methods they use under relaxation factors so heavily! In some sense the stability provided by the under relaxation scheme is related to the same CFL condition.

Re:Limits on simulation.

[jstults]

In some sense the stability provided by the under relaxation scheme is related to the same CFL condition.

Not at all. You use under-relaxation because you aren't in the 'ball of convergence' for your Newton method, so you don't take the full step. The CFL condition is a condition on stability of the discretization. Two completely different things.

Re:Limits on simulation.

[dontmakemethink]

The CFL condition that limits the maximum time step one can take shows no sign of relenting.

You mean they're STILL punting on 3rd down?!

Re:Limits on simulation.

[spirito]

You are right, fluid dynamics simulations parallelize beautifully, but once you start increasing the number of cores the communication between machines will slow things down. And you can have implicit, time accurate, temporal schemes for which the stability condition is (theoretically) CFL less than infinity. But it's clear that if you want to resolve turbulence time scales (and length scales) on a complicated case with relatively high Reynolds number a 100 million processor machine may not be enough.

How many problems can these systems really solve?

[wondi]

All this effort at creating parallel computing ends up solving very few problems. HPC has been struggling with parallelism for decades, and no easy solutions found yet. Note that these computers are aimed at solving a particular problem (e.g. modeling weather) and not at being a vehicle to quickly solve any problem. When the comparable multi-processing capacity is in your cell phone, what are you going to do with it?

Why 100 million processors?

[140Mandak262Jamuna]

Technically, shouldn't 640K processors be enough for every one?

Re:Why 100 million processors?

[Yvan256]

It is, if we're talking about cloud processors for running vaporware.

Re:Why 100 million processors?

[aztracker1]

VMWare is now renaming its' company to "CloudWare" which will be optimised for use with Intel's new Vapor® processor line.

Portland, Ore?

[Anonymous Coward]

What's "Portland, Ore"?

Oh, you mean "Portland, Oregon"? This is a website. It isn't fucking Twitter or SMS, how hard is it to write three more letters?

We know that Slashdot is U.S.A.-centric so we'll forgive the missing "Portland, Oregon, U.S.A." part, but for crying out loud, at least write the whole state name.

Re:Portland, Ore?

[Anonymous Coward]

Why 3 letters? the standard abbreviation for a state is 2 letters, in this case OR

Re:Portland, Ore?

[mandark1967]

how hard is it to write three more letters?

If I had to hazard a guess, I'd say it is exactly 3 letters harder to write 3 more letters.

Unless of course, you bold the last 3 letters, as you've done...then you have the html code to type so it ends up being like 10 letters more difficult.

Then again...it depends on whether or not you mean 3 "additional" letters or the phrase "3 more letters." because, in that case, it's like 13...even more if you bold some of them...

Re:Portland, Ore?

[Anonymous Coward]

FYI, <b> </b> is deprecated. You should take the habit of using <strong> </strong> ASAP.

Oink, oink

[Animats]

The exascale systems will be needed for high-resolution climate models, bio energy products and smart grid development as well as fusion energy design.

Sounds like a pork program. What are "bio energy products", anyway. Ethanol? Supercomputer proposals seem to come with whatever buzzword is hot this year.

It's striking how few supercomputers are sold to commercial companies. Even the military doesn't use them much any more.

Re:Oink, oink

[Anonymous Coward]

Sounds like you're full of bullshit.

Re:Oink, oink

[David Greene]

Sounds like a pork program. What are "bio energy products", anyway. Ethanol?

I'm no expert on this, but I would guess the idea is to use the processing power to model different kinds of molecular manipulation to see what kind of energy density we can get out of manufactured biological goo. Combustion modeling is a common problem solved by HPC systems. Or maybe we can expore how to use bacteria created to process waste and give off energy as a byproduct. I don't know, the possibilities are endless.

It's striking how few supercomputers are sold to commercial companies. Even the military doesn't use them much any more.

Define "supercomputer." Sony uses them. So does Boeing. The auto industry uses clusters to model crashes, but I believe that's more limited by the design of the off-the-shelf software than anything. They could certainly run on supercomputer-class machines if the vendors ported them.

And the military uses them a lot. Much of the DOE research done on these machines is probably defense-driven.

Re:Oink, oink

[vtcodger]

Isn't quantum computing supposed to solve all these problems without need for a zillion cores? Or have a latched onto the wrong panacea here?

Re:Oink, oink

[T Murphy]

These systems cost a lot- it might take buzzwords to get politicians to buy into them and fund these sorts of projects. Even so, many energy projects are important to pour more research into, even if such projects often get watered down to a single misleading buzzword.

Re:Oink, oink

[NewbieProgrammerMan]

It's striking how few supercomputers are sold to commercial companies.

I'm sure that in the early 20th century somebody was saying, "It's striking how few airplanes are sold to commercial companies," and going on to draw the conclusion that government spending on aircraft was a pork program. (And I'm sure there was some pure pork spending involved, but 100 years later, the overall effect of that kind of spending lets us use airplanes for things that would have been unthinkably expensive when people started spending money on them).

Today's supercomputer is the next decade's mid-range workstation. (Yes, I know, duh.)

Re:Oink, oink

[DriedClexler]

Well, OBVIOUSLY it's to provide a basis for those cringe-inducing IBM ads where they talk in really dumbed-down terms about how they're going to make a "smarter planet" by making everything run "smarter" because IBM is going to throw a shitload of supercomputing power at it, and have smart- or foreign-sounding people talk about how great it is. (Nevermind that e.g. the alleged traffic congestion reduction was due to peak-load pricing, not to the ability to crunch numbers at supercomputer speeds.)

Example.

Get with the program, man! ;-)

AMD vs Intel

[teko_teko]

It's interesting that 4 of top 5 supercomputers are running AMD, while 402 of the Top500 are running Intel.

What's the cause of this? Value? Energy-saving? Performance?

Re:AMD vs Intel

[Eharley]

I believe AMD was the first mass market CPU to include an on-board memory controller.

Re:AMD vs Intel

[jamzo]

It's probably because of HyperTransport and floating point computation performance. I think HyperTransport made it easier for supercomputer vendors like cray to build better interconnects and traditionally opterons was a bit better at floating point ops. Also, opterons have 64Kb L1 caches where I think comparable Intel processors had 32Kb L1 caches. But this was all a couple of years ago ... the next generation of the fastest supercomputers will probably be Intel based.

Re:AMD vs Intel

[Anonymous Coward]

Not energy saving, since the Cell based ones are clearly on top in performance/Watt. Note that number 2 is actually an AMD/Cell hybrid where the majority of flops are provided by Cell. It has 2/3 the performance of number 1 but 1/3 the power consumption. And the x86 based that has the same power consumption has half the flops.

I'm curious to see what it will be in 1 year, since Power7 might be a serious contender. (Power6 has been IBM's Pentium IV, very high clock speed for limited performance and performance/Watt, but first Power7 early benchmarks look much better).

Power7 blades will be more expensive but the Power7 also has in theory more memory bandwidth, scales better to a larger number of threads per memory coherency domain, and may have much better performance/Watt. In this case, the larger upfront costs may not be decisive (especially when the difference is counted in MW, evacuating them has an impact on the infrastructure).

Re:AMD vs Intel

[confused one]

I'd be guessing but here are three possible reasons AMD might be in that place:
1.) Value, ie. lower cost per processor
2.) Opteron has built in straight forward 4-way and 8-way multiprocessor connectivity, Xeon was limited to 2-way connectivity without extra bridge hardware, until recently.
3.) Opteron has higher memory bandwidth than P4 or Core 2 arch.

Re:AMD vs Intel

[hattig]

Easy CPU upgrades because the socket interface stay the same.

Some of those supoercomputers might have gone from dual-core 2GHz Opteron K8s through quad-core Opteron K10s to these new sexa-core Opteron K10.5s with only the need to change the CPUs and the memory.

Or possibly if the upgrades were done at a board level, HyperTransport has remained compatible, so your new board of 24 cores just slots into your expensive, custom, HyperTransport-based back-end. To switch to Intel would require designing a QPI-based back-end.

Of course Magny-Cours and Bulldozer will use the G34 socket, so that's not a plug-in and go upgrade when they come out in 2010 and 2011 respectively. But it will be a stable platform for several years itself, and thus be attractive.

Re:AMD vs Intel

[Anonymous Coward]

As much as I think AMD are a bunch of crybabies, they made a wise decision early on with Hypertransport and with the IMC. They had a solution that worked well from the desktop on up to HPC, while Intel mainly targeted the desktop up through 2P systems. They certainly had solutions for 4P and above, but none were as elegant as the Opterons were for that space.

Note that with Nehalem (and the EX version) this will change and Opteron is no longer compelling in, well, any space other than on a price/performance basis possibly.

Re:AMD vs Intel

[jd]

It was the first mass-market ix86-clone CPU to include an on-board memory controller. The Transputer had a perfectly good on-board memory controller 25 years ago. It was 64-bit, too.

Re:AMD vs Intel

[jd]

It's doubtful this is a factor, but AMD's HyperTransport 3 has higher bandwidth than Intel's PCI Express 2.1.

Why build this monstrosity?

[140Mandak262Jamuna]

We know what answer it is going to give. 42. Save the money.

Re:Why build this monstrosity?

[thewils]

That's the answer though. They're building this thing to find out what the question was.

Re:Why build this monstrosity?

[Anonymous Coward]

Wouldn't that cause the Universe to crash with a Stackoverflow, with the computer generating the plan for a world to compute the question and the world building a computer to solve the question and the computer #CONNECTION LOST NO CARRIER

The Jaguar?

[Yvan256]

The Jaguar is capable of a peak performance of 2.3 petaflops.

The first Jaguar was a single megaflop.

Re:The Jaguar?

[threephaseboy]

But did it leak oil?

Re:The Jaguar?

[Yvan256]

Well, okay, the second Jaguar (cue reply about the Atari Jaguar not being the second commercial product called Jaguar).

Also, the person who modded my post above "interesting" is either on crack or I was right (by luck) about the Jaguar having 1 megaflop of computing power.

My post, however, was that the Atari Jaguar was a mega-flop, i.e. its sales were abysmal, support non-existent, etc.

Partly a software problem. Erlang?

[mcrbids]

We're still at the point where unthreaded languages (like PHP) are still viable. For example, we use PHP in a complex, multi-server, multi-core cluster, and it's "share nothing" approach scales quite nicely, in that having more and more users hitting the systemm on separate servers doesn't really cause a problem, since there's virtually no cross-communication going on.

But there's a scalability limit in what you can do "PER PROCESS". There are some very processor intensive functions that simply take a while to do (such as rendering a 100 page report, then converting to PDF) and there's currently no way to spread the load in PHP beyond a single core.

At the other extreme, we have almost the same problem - having such a large number of cores that resources commonly shared among threads and processes is really no longer feasible.

Languages like Erlang have a "shared nothing" approach, but not at the process/thread level, but at the function level. Individual functions within a process are themselves "share nothing" and thus can easily scale across multiple cores, processors, and servers in a networked cluster. (at least, this is the theory)

So how 'bout it, folks? Where are the benchmarks showing how languages DESIGNED to take advantage of parallel processors and clusters actually scale up in the real world? Is Erlang the cat's meow when discussing systems of this scale?

I'm not expecting to see my example process (100 page PDF reports) scale up smoothly to 250,000 cores, but I sure would like to see it scale up smoothly to a dozen or two!

Re:Partly a software problem. Erlang?

[Anonymous Coward]

What on Earth? You're bringing PHP and "rendering PDF reports" into a discussion about HPC? And you propose Erlang as some kind of solution? Nobody doing HPC is using Erlang. As usual for Slashdot, you have absolutely no idea what you're talking about.

Re:Partly a software problem. Erlang?

[vlm]

I'm not expecting to see my example process (100 page PDF reports) scale up smoothly to 250,000 cores, but I sure would like to see it scale up smoothly to a dozen or two!

Well, that's not very hard. Split the job like the ray tracers do, into 250K little parts of the 100 page report, have each core individually render its little bit, then mush all the rendered outputs together.

You could do this now, more or less off the shelf, by separating your raw data into 100 raw input files, one for each page, then have 100 machines or cores or whatever render each separate page, then a big run of pdfjoin to turn 100 single page pdfs into one 100 page pdf.

Re:Partly a software problem. Erlang?

[mcrbids]

You're bringing PHP and "rendering PDF reports" into a discussion about HPC?

Yes. Because in a few years, hardware comparable to what is now "HPC" will be routine. We've already jumped from unicore, uniprocessor servers to almost a hundred cores. Just a decade or so, a 100-core computing cluster was HPC, even if not near the "top 500".

And you propose Erlang as some kind of solution? Nobody doing HPC is using Erlang.

But... Erlang was designed for scalability! It was DESIGNED to smoothly scale from a unicore to multicore to LOBOS style computing. If Erlang isn't a "player" in the HPC space, why the !@# not? And if it's not a player, what is, and what do I need to do to transition to it over the next decade or so?

As usual for Slashdot, you have absolutely no idea what you're talking about.

Since you posted this on Slashdot... (I'll put the mirror away)

Re:Partly a software problem. Erlang?

[mcrbids]

Have you *tried* to do this with PHP?

The only meaningful way to do this is to make an HTTP call to self, which is expensive, or write it in C or something and shell out, which makes doing it cross-cluster painful without adding lots of overhead and complexity.

PHP just isn't the "right" tool for this type of job. Projects like PHP/Erlang might mitigate this, somewhat. Thus my probing in re: HPC and Erlang, since massively multicore is obviously the future.

In other news: it's surprisingly difficult to perform useful neurosurgery with a hammer.

Re:Partly a software problem. Erlang?

[David Greene]

If Erlang isn't a "player" in the HPC space, why the !@# not?

Because it's not Fortran, C or C++. Really. That's the breaks. It's very hard to introduce a new language into a mature market.

It's functional nature also makes Erlang less attractive. While the theoretical benefits are great, various practical issues often get in the way. For example, how does one write a robust, efficient database in a functional language with single assignment? Finally, its dynamism comes at a cost that is often unacceptable to HPC users. Message passing has similar consequences.

And if it's not a player, what is, and what do I need to do to transition to it over the next decade or so?

Probably OpenMP, Co-Array Fortran and UPC.

Re:Partly a software problem. Erlang?

[shmlco]

"... by separating your raw data into 100 raw input files, one for each page..."

Except that most reports have totals and breaks and subtotals that wouldn't work across such pagination. In fact, you'd practically have to paginate the entire report first to determine what elements to assign to each page.

Re:Partly a software problem. Erlang?

[dlapine]

Well, the increase in numbers of cores will surely migrate down to the desktop level. Don't both Nvidia and ATI claim to provide graphics cards with hundreds of cores now? All available to the casual user.

He's got a point, though. Although the current HPC market is mature, and very sensitive to performance, the upcoming market for desktops with 100+ cores (call it hecto-scale) will care less about efficiency, and more about providing an easy way for the users to actually utilize all those cores. Even if the erlang language isn't want today's HPC users want, something like it will be useful for users on those hecto-scale desktops. I want to be able to take advantage of all those cores, but I sure don't want to write MPI or any other of today's options for massively parallel computing.

Maybe OpenCL or CUDA will spawn off some interpretive language that will be useful for us non-parallel programmers. On a system with 100+ cores, a language can afford to be less efficient for scripting or simple computing, but still find a use for all that processing power. We've certainly been willing to accept a loss of efficiency versus an ease of programming once the systems get powerful enough (look at the original assembly programming for the 8 bit home systems and compare it to say, java and an IDE).

Re:Partly a software problem. Erlang?

[Anonymous Coward]

What on Earth? You're bringing PHP and "rendering PDF reports" into a discussion about HPC? And you propose Erlang as some kind of solution? Nobody doing HPC is using Erlang. As usual for Slashdot, you have absolutely no idea what you're talking about.

Perhaps rather slating people you could explain why what he said didn't apply. Just because you may or may not think you know more than him doesn't give you the right to be a) rude to him and b) slate slashdot readers in general.

creators' big flash coming way before 2018

[Anonymous Coward]

that will settle who 'owns' what forever.

it's newclear powered, user friendly, & completely bug free, as well free to use, forever.

Windows 2018

[smitty777]

Maybe this thing will have enough power to run Windows by 2018??

Re:Windows 2018

[Anonymous Coward]

Windows has the requirement of SOTA +1, so that it can be used for 5 years into the future per its upgrade path. This also means that it isnt worth buying the new windows until you have a new machine and SP1 has been released.

Re:How many problems can these systems really solv

[Dgtl_ _Phoenix]

When the comparable multi-processing capacity is in your cell phone, what are you going to do with it?

Stream high definition porn... duh.

human brain

[simoncpu was here]

How many cores do we need to simulate a human brain?

Re:human brain

[Dgtl_ _Phoenix]

Not sure about the number of cores, a number of experts say that around 20 petaflop should do it. We should see computers capable of doing this by the end of the decade. Of course creating the AI or brain scans necessary to accomplish this is going to be the more challenging problem. What will be fantastic about simulated brains is that their neurons will be significantly faster than standard human neurons. This means that your simulated brain can produce orders of magnitude more work despite being no smarter.

Re:human brain

[Anonymous Coward]

Only 1, so long as the simulation segfaults.

Re:human brain

[CannonballHead]

Define "simulate" in this context. Processing power? Creativity? Originality? Ingenuity? I didn't think any number of cores could "cause" creativity... aside from a "brute force" method. Try-every-possibility-and-see-if-one-works.

Re:human brain

[iprefermuffins]

One for each neuron should do nicely...

Re:human brain

[Anonymous Coward]

One?
Once you have enough knowledge about how the brain works, it shouldn't matter if it gets simulated in hardware or software.

Re:human brain

[Anonymous Coward]

Let's come up with an algorithm to simulate the human brain before we think about the amount of cores it requires.

Re:human brain

[ari_j]

Which human brain? I can simulate the rational thought processes of 90% of humans with one vacuum tube.

Re:human brain

[AmigaMMC]

About a Gazibillion

Re:human brain

[im_thatoneguy]

It's a lot easier to test algorithms if it doesn't take 10,000 hours to simulate a fraction of one second. Even if you perfect replicated the algorithm for a human brain what would it learn in 1/100th of a second simulated? It's really really hard to adjust simulation algorithms when you can only make a change every year. And that is assuming we're just tweaking a model that is close. It's going to take a system fast enough to simulate weeks or months in a a few hours before we can start really tackling the problem.

Re:human brain

[argoth]

This means that your simulated brain can produce orders of magnitude more work despite being no smarter.

I've had several managers with the same ability.

Re:human brain

[Anonymous Coward]

One http://en.wikipedia.org/wiki/Turing_completeness

The issue is understanding of the brain far more so than computing power.

Re:human brain

[caywen]

I have this feeling that simulating the human brain ends up being practically meaningless without simulating the rest of the body. And, having some meaningful, long-term interaction with either the real world or an incredibly sophisticated simulation. I'm assuming one would "grow" the brain in the simulation rather than, say, capturing the full chemical state of every neuron in a living person's brain (which sounds pretty impossible to me).

Re:human brain

[jwhitener]

You need about 10^14, or 100 teraflops, or 100 trillion calculations per second, in parallel. That's the best guess for the raw computation power of the brain.

Now thats just matching it, with no software and little understanding of many parts of the brain. Some people suggest that the brain has other stuff going on that might be very hard to figure out from current mapping/scanning techniques. I forget his name, but one researcher suspects that there are quantum effects, maybe even quantum level calculations, happening inside tiny parts of the brain that are affecting the outcomes. And other researchers are looking at the effect of signal strength in neuron firing. So not only are there 100 billion neurons each connected in a network performing 100 trillion neuron firings per second, but those firings are of various strengths which means that it is conveying more information than just "fired" or "not fired".

I think thats how I remember a summary from some AI article. It boils down to the brain most likely having orders of magnitude more calculations per second than Hans Morvec estimated. And your computer running the simulation is going to have all the overhead of creating the virtual biology. That is, unless we can really understand what is important to the brain and what is not. You can simulate an analog signal in a digital computer, but there is overhead involved. 100 trillion analog signals means your computer simulation is going to be doing way more than 100 trillion calculations.

Re:human brain

[lennier]

"I can simulate the rational thought processes of 90% of humans with one vacuum tube."

And you get a much warmer sound that way, too.

Re:human brain

[Surt]

That's a bit low. When I last did work in the field, low accuracy neuron simulations were using about a 486 (1 megaflop) to simulate one neuron in 1:100 of real time. The brain has about 10^11 neurons, and mega to pet is 10^9, so that's 10000 petaflop for low accuracy. High accuracy (probably necessary for something that is going to simulate a brain) will need 10x as much or more. This is all assuming that none of the interactions scales poorly.

Re:human brain

[haxor.dk]

Slightly over 9000.

Re:human brain

[ari_j]

I was thinking more a 5Y3. No switching and the only effect on character is a whole lot of sag.

Re:human brain

[Anonymous Coward]

the amount of transistors is the least of our problems. we need a complex organization of those units in a way that we cannot fathom, and are not likely to be able to any time soon.

Windows Vista

[Anonymous Coward]

The question is: will it be enough to run Aero?

Re:Windows Vista

[MickyTheIdiot]

no

So far to go!

[Anonymous Coward]

The fastest system only has 224k Cores? Oh, dear. We definitely need bigger systems, then.
And I suppose Deep Thought has nothing to worry about yet, either. Yet. :D
(The fictional version, that is. The "real" one has already been outdone by Rybka.)

I can visualize a third life

[ub3r n3u7r4l1st]

Yup. 100-million core driven Second-Life server, like the Matrix.

That's a lot

[Anonymous Coward]

Wow, not just one million-core supercomputer, but 100 of them?

Speaking of heat

[ArbitraryDescriptor]

I am currently accepting investors to help build a one billion core supercomputer to create high resolution climate models that take into account the waste heat from a 100 million core supercomputer making a high resolution climate model.

(Seriously, how much heat is that thing going to put out?)

Re:Speaking of heat

[Anonymous Coward]

I was reading on wikipedia that MSoft sold 400 million copies of windows XP.
This doesn't incude the pirated ones. Thus this means that there are 400 million PCs in the world. If everyone had the newest PC with 6 core athlons, there would be 2400 million cpus in the world. So we've already met the 100 mil core limit time 24!!! Thus, exaflop has already been reached. What's the next one one? Hmmm, TSA can decode all the emails everyone sends out, including the millions upon millions of spam all the faster..... Thus the govs supercomputer is sure one glorified spam checker.

Re:Speaking of heat

[blueg3]

(Seriously, how much heat is that thing going to put out?)

As much energy as it consumes. For climate models, though, direct waste heat production is negligible compared to climatological effects (e.g., CO2).

Re:Speaking of heat

[DarthVain]

Ah Heisenberg you crafty devil!

Re:Speaking of heat

[AmigaMMC]

I was reading on wikipedia that MSoft sold 400 million copies of windows XP.
This doesn't incude the pirated ones.

.

Obviously it doesn't or you'd be saying that Microsoft has sold pirated copies of Windows ;-)

Re:Speaking of heat

[Anonymous Coward]

Just because 400 million copies were sold does not mean there are 400 million PC's. Take for example a person who bought a computer with XP home and later decided they wanted XP Pro, so bought it and installed in on the same computer already running XP Home. That counts as 2 sales, but only one computer.

Re:Speaking of heat

[jd]

It depends on what sort of cores are used. Most real problems are not SIMD but MIMD. You can turn a MIMD problem into a very large set of SIMD problems, though, where each one of those SIMD problems is actually extremely narrow and does not require a completely general-purpose core.

What you end up with is a hybrid of highly specialized cores, each very tightly tuned to some specific sub-category of problem. Since it's tuned, it can be ultra-RISC (it might not even need to be Turing Complete in some cases).

Since the critical path determines deadlines, and most processes will be OFF the critical path, you can also have asynchronous cores. The University of Manchester's AMULET group has been working on asynchronous computing for a while and has some nice technology, including open-source tools for making asynchronous chips.

Switching is another big heat-generator, since extreme computing like this really needs either a butterfly topology or a hypercube topology. In either case, IIRC, the number of switches goes up with the power of the number of nodes. At 100 million nodes, you'd be looking at tens of quadrillions of switches. Even if each switch is relatively lightweight on power, the sheer number makes the total consumption significant.

However, it's not all bad. Any collective operation can be delivered via NACK-oriented reliable multicast, so it's a single-shot delivery to as many nodes as you like.

Since supercomputing IS all about critical paths and deadlines, it should be possible to use delay tolerant networking protocols and have switches deliver as much in bulk as possible to minimize processing needed except on the periphery of the network. So long as packets get to their destinations by the deadlines, it makes no difference to the nodes whether they get the data then block on a barrier operation or whether they get the data just as the barrier operation is due to finish.

So there's lots of ways to cut down on the heat on the hardware side. What about the software side? Well, competent programmers are often a good start. A million cores doesn't mean it's ok to be a million times less efficient.

Code should be tightly written and customized for the specific hardware (including network) characteristics. It should then be profiled, baked, broiled and fried until light and crispy. The programmers should then refactor the code until it's done right.

The theoretical minimum heat output from a hybrid asynchronous computer is the heat you'd generate if one core is one processing element with enough memory to perform its task. You can't get any more RISC than a single instruction. The theoretical minimum heat output from a deliberately delaying network is the heat you'd generate in queueing and bulk-delivering data such that nodes are either off or busy, they're never idle on the offchance of data coming soon.

Idle nodes are the really bad nodes. If a node cannot be used right then and is off, it consumes nothing. A node is most likely to be idle if there is simply no way of telling if it'll be assigned work or not.

The next-worst are nodes that do things too fast. Doing things too fast means the memory has to stay powered to preserve the results but nothing can consume the results any time soon. Power consumed is not directly proportional to processor speed. It's not linear. If you detune the processor such that it always has the data ready just in time to be consumed, you will always use less power than storing.

The upshot of this is your input buffer wants to be very large (so you can bulk-receive) but the output buffer wants to be comparatively small (because you are generating results when they're needed and no sooner).

If you're really smart, you'd make the input buffer high-bandwidth and the output buffer high-speed. If you're generating just-in-time, the last thing you want is for the output to be slow in being forwarded on. On the other hand, data will be dropped into the input far faster than it can be consumed (since it's a bulk delive

Re:Speaking of heat

[turing_m]

(Seriously, how much heat is that thing going to put out?)

Your joke was funny. To answer your serious question, probably less heat than you might think. For example, the i5-750 with the so-called "turbo boost" uses one or two cores running fast and taking up all the power when that is called for, or potentially all 4 cores running within the same power envelope if there is something that can use 4 cores. That's 4 cores doing much more work than 1, and using the same amount of power. More cores running slower must continue scaling, otherwise they wouldn't be doing it.

Re:Speaking of heat

[dontmakemethink]

I wouldn't even power up a 100M core box that couldn't figure out what to do with its own heat.

"Sort that heat out or I'll send you into a perpetual loop re-transcoding 'Iceland's Got Talent' re-runs. You hear me?!"

Re:Speaking of heat

[Ultracrepidarian]

Actually, the known universe is just a part of a massive simulation.

Re:Speaking of heat

[hesaigo999ca]

>Seriously, how much heat is that thing going to put out
Enough to burn a hole in the ozone layer....or melt the arctic polar cap.
Seriously though, if we could also develop something to harness the neat this thing will give off, it would be a double whammy!

Re:How many problems can these systems really solv

[David Greene]

Note that these computers are aimed at solving a particular problem (e.g. modeling weather) and not at being a vehicle to quickly solve any problem.

That's not entirely accurate. HPC systems are designed to solve a class of problems. That's not the same thing as a "particular" problem. Jaguar has, in fact, solved many different problems, including fluid flow, weather, nuclear fusion and supernova modeling. It's not going to run Word any faster than your PC but that's not what you buy a supercomputer to do.

synergy

[Anonymous Coward]

They should obviously start working with the Mandlebulb people..

Processing power

[Wowsers]

Is this going to be the new processor requirement for running Flash in a web browser?

Re:Processing power

[jd]

No. It's the new processor requirement for Windows 8. Running the web browser requires an additional 10 million cores. If you want Flash as well, better make it 25 million extra.

Re:How many problems can these systems really solv

[Anonymous Coward]

These computers will be good for solving problems involving lots of independent operations. A processor can process one operation at a time, but since these operations do not depend on each other, the operations can be sent to several processors at once. Imagine a big foreach loop, for example.

Or maybe someone will want to have 100 million Google Chrome tabs open.

I'm personally imagining putting dnetc on one of these things.

Sure

[Stan Vassilev]

I'm still waiting for that 10GHz Pentium Intel promised for 2004.

Re:Sure

[imamac]

http://www.newegg.com/Product/Product.aspx?Item=N82E16819115043

Re:How many problems can these systems really solv

[2obvious4u]

Parallel computing is great for solving NP-Complete problems. If you have enough cores for every possible solution you can have all possible paths process at the same time and compare the results.

Re:How many problems can these systems really solv

[Again]

That's not entirely accurate. HPC systems are designed to solve a class of problems. That's not the same thing as a "particular" problem. Jaguar has, in fact, solved many different problems, including fluid flow, weather, nuclear fusion and supernova modeling. It's not going to run Word any faster than your PC but that's not what you buy a supercomputer to do.

So you're saying that OpenOffice would still take forever to start.

Coming By 2012

[Anonymous Coward]

Republican reign for the next 1000 years after Obama fails.

Yours In Moscow,
Kilgore T.

Re:Coming By 2012

[Surt]

Not much chance of him failing at this point. Unless we hit a second dip in this recession, Obama will successfully claim his policies are what turned things around during the 2012 campaign, and there will be no gainsaying it.

Sir, all 1 million cores have failed..

[Anonymous Coward]

if only they had built 1,000,001!

How about reconfigurable computing instead?

[Anonymous Coward]

...Such as, say, FPGAs (www.maxeler.com) or GPUs (www.ati.com)
One such accelerator card can replace ~ 100 cores for common applications such as finite differences or MonteCarlo ...
Hence you would need "only" a million blades.

One Hundred Billion Cores

[Anonymous Coward]

Made by sharks with frikkin lasers on their heads.

Is it just me or does this news message just shout Dr. Evil?

You sign your posts manually?

[Anonymous Coward]

What's it like to be a primitive caveape in the modern world?

20 Megawatt power supply...

[tomhath]

IBM's design goal for an exascale system is to limit it to 20 megawatts of power ,,,

Just keeping that sucker cooled will contribute to global warming. I hope they're going to use all that waste heat for something.

Colon Blow

[Beelzebud]

To put this into perspective, it would take over 4 and a half million bowls of Super Colon Blow to equal the computation power of just 1 of these things!

stupid

[Anonymous Coward]

"But the problems involved in reaching exaflop scale go well beyond Moore's Law."

  The above quote shows quite well that the writer doesn't understand what Moore's Law is about.

Are all your cores used on your desktop?

[Ilgaz]

Check Wiki about "thinking machines", "transputer" and if you have more than 1 CPU/Core, launch a game and see if all cores used effectively without needing massive additional work from game publisher.

Technology is primitive, even a billion processor machine doesn't save it from being primitive. It is the software at least.

Ok, Ok, we'll be more specific...

[raftpeople]

"Portland, Oregon, U.S.A., Earth, Milky Way, Cluster TXH-170718, Universe 01 (we think)"

Re:Ok, Ok, we'll be more specific...

[AmigaMMC]

"Portland, Oregon, U.S.A., Earth, Milky Way, Cluster TXH-170718, Universe 01 (we think)"

Quite presumptious on your part. It's Universe 84F32G Signed Dr. Zoidberg

Re:Ok, Ok, we'll be more specific...

[raftpeople]

Is that why I'm not getting my mail?

Re:How many problems can these systems really solv

[T Murphy]

I'm always wary of making an infamous "50 MB of memory is all you'll ever need" type of claim, so I like to believe that we'll figure out how to use greater processing power by the time it gets here. We haven't had too much trouble with that so far. As far as actual use, if we ever get products like Morph (http://www.youtube.com/watch?v=IX-gTobCJHs), there might be a need for massively parallel processing. At the very least, such computing power would likely be needed to make such products.

address lines

[conspirator57]

if speed is the goal, then it needs to be a power of two number of cores so that you don't have to implement logic checking for a valid core address. That logic would eat performance from every action performed by the machine. So, until you develop affordable decimal logic hardware implementations that can scale in size the way the binary logic does, we're gonna keep making computers that work fast the way we do now and it's gonna involve powers of 2. And get off my lawn.

BILL LIVES!

[TiggertheMad]

just had an ugly thought...."Windows17 for PC (Personal Cloud)"

100 million-core supercomputers?

[Tetsujin]

Come on, that's just silly. I can understand why we might a few million-core supercomputers, but who would need 100 of them?

Re:100 million-core supercomputers?

[jd]

Depends on how many turn up to the LAN party. C'mon, have you SEEN the new Quake 8 graphics?

Re:100 million-core supercomputers?

[Tetsujin]

Depends on how many turn up to the LAN party. C'mon, have you SEEN the new Quake 8 graphics?

Does that mean they've finally gotten the hang of optimizing game code for parallelism?

Re:100 million-core supercomputers?

[jd]

Well, not exactly. They're using a modpack for The Sims to develop virtual programmers who can write parallel game code.

Interconnect?

[nokiator]

Scaling the number of cores to 100 Million by 2018: A side effect of Moore's Law.

Low latency, high bandwidth interconnect that can mesh 100 Million cores: The Next Big Problem in computer architecture.

Re:Interconnect?

[FishOuttaWater]

That's a big issue, yes. Also fault tolerance. You gotta map out the dead processors and communications nodes (in real time?), move the load around to compensate, and log the problems for maintenance.

"now gathering in Portland, Ore"

[Tetsujin]

Can I just say... FUCK YES. Thank you!

As someone who grew up in the Portland (Maine) area it annoys me to no end when people talk about things in "Portland" and neglect to disambiguate - especially when they're talking about the other Portland. :)

Re:"now gathering in Portland, Ore"

[Bitmanhome]

And I'm glad that they clarified that they're not talking about the Portland operating room. Although I still don't know what's so special about Portland rocks.

Re:How many problems can these systems really solv

[von_rick]

Yes. So by extending it to a million core machine, OpenOffice would take million x forever* to load if one instance is opened per core.

*Forever = two seconds after a mouse click.

100 million cores

[joetomato]

To which Oak Ridge National Laboratory replied "Fuck everything, we're doing 500 million cores."

Too little too late!

[syousef]

I take it THIS is a machine that might run Vista well. Too late SP3 aka Windows 7 is out.

Re:How many problems can these systems really solv

[shmlco]

Stream it? With that much processing power it should be able to create it on the spot: "Computer, let's start today's scenario with Angelina Jolie surrounded by...."

Actually...

[Anonymous Coward]

If you think massively parallel supercomputers are like throwing logs on a fire, you should also think rocket motors are like throwing more fuel on a fire. Both are abstractly about the same goal of condensing more 'reaction' into a single motor, and both are devilishly complex in all the details which distinguish them from scaling a camp fire to a bonfire. Scaling up manufacturing and delivery of the fuel units to all pack consistently into one system which can be controlled in practice, while avoiding issues like uncontrolled resonance, heat dissipation...

Windows 15

[KaoticEvil]

System Requirements
100Million Core CPU...
5 Gigabytes RAM
1 Petabyte of hard drive space...

Yeah, I can see it..

Re:Windows 15

[sznupi]

50 bytes of RAM per CPU?...

Funny, fusion...

[bryan1945]

Probably will need a fusion plant to power and cool the thing. But still sounds awesome. They briefly mention data/memory flow issues, but don't really address it. It is getting to the point where data flow will be as important as processing power, especially as you have escalating processors. You can run as many operations as you want, but if it can't be delivered somewhere useful, then they are wasted. I am also very interested on how the overhead will be managed when this many processors are involved. Multi-processors are not quite 2x (or 4x, 8x, etc) than just one processor due overhead, and even a really specialized scaled OS & I/O system won't be able to overcome this many processors.

Now for some fun:
It could probably power real time rendering of a Beowulf cluster of Natalie Portmans in grits while making us submit "All Our Base" to our new "Insert-Here Overlords". (Did I miss any?)

Re:Funny, fusion...

[haruchai]

Perhaps they should name this machine the Catch-22 - we need something this powerful to help solve
the problems of nuclear fusion - which would be needed to power this box.

Re:Funny, fusion...

[bryan1945]

I likey! I'm sad I didn't think of this myself.

Re:Funny, fusion...

[Anonymous Coward]

It could probably power real time rendering of a Beowulf cluster of Natalie Portmans in grits while making us submit "All Our Base" to our new "Insert-Here Overlords". (Did I miss any?)

In Soviet Russia, warn out memes miss you!

First Use of New Machine

[khelms]

Q: Is there a God?
A: There is now!

Perspective

[Anonymous Coward]

Let’s put this into perspective: Intel’s next generation of chips are estimated to have ~2 billion transistors (release in 2010 or 2011). Moore's Law would predict that chips would have ~32 billion transistors in 2018. That’s an estimated 3.2E+18 transistors spread over the 100-million cores

That’s 32,000,000 times the number of neurons in the human brain.

I'd like to see. . .

[Fantastic Lad]

Department of energy?

Mapping weather systems?

Cracking high bit encryption schemes? Listening to every phone call happening on the planet and mapping social patterns?

BORING!

No, I want to see a 100 million core supercomputer render one of those 3D "Mandelbulbs" and let me do some real-time exploring with a VR helmet.

Now THAT would be a worthy use for such resources!

That and being able to grow virtual beings from DNA samples.

-FL

Obligatory

[mhajicek]

All your cores are belong to us!

Re:Obligatory

[Anonymous Coward]

If by "Obligatory" you meant "Retarded", then, well, yeah, knock yourself out, kid.

what do we do

[confused one]

when these exascale systems start asking questions and/or making demands?

Re:How many problems can these systems really solv

[1729]

Parallel computing is great for solving NP-Complete problems. If you have enough cores for every possible solution you can have all possible paths process at the same time and compare the results.

That's tough to manage when the number possible paths grows exponentially with respect to the input size.

Boom, boom

[AlpineR]

It's striking how few supercomputers are sold to commercial companies. Even the military doesn't use them much any more.

Oak Ridge National Laboratory, home to the world's fastest supercomputer, does a lot of work for national security. At the labs housing the top ten supercomputers, at least five do weapons and defense research. And that's just what the public knows about. I would be shocked if there weren't similar supercomputers working on intelligence and classified projects.

Even if the computers aren't stamped with "U.S. Army", the military does indeed use many of them. The wonderful side effect of their push to simulate things like aging nuclear weapons is that it helps develop the technology for peacetime purposes like renewable energy and pharmaceuticals.

Re:How many problems can these systems really solv

[jstults]

How many problems can these systems really solve?

Well, only the ones where you need to conserve mass, momentum and energy; pretty niche market really...

Add as many processors as you want....

[Ozlanthos]

I am sure Stalker SOC will still crash.

-Oz

$327.67, $83,886.07

[tepples]

But even when choosing a data type for money, an app designer still needs to know base 2 to find the appropriate type in MySQL capable of holding amounts up to $327.67 (smallint), $83,886.07 (mediumint), over $21 million (integer), or more money than the gross world product (bigint).

Re:$327.67, $83,886.07

[aztracker1]

try Decimal(65,8)

Re:$327.67, $83,886.07

[tepples]

try Decimal

Provided that the language in which your application is implemented supports decimal types. If you know your fractions will all have the same denominator (like 100 for money), it may be more efficient just to use integer types. Case in point: An older version of the software that runs my employer's business used Python's decimal.Decimal type, but it was dog slow, so we switched to integers representing U.S. cents.

Re:How many problems can these systems really solv

[Surt]

A 100 million cores supercomputer will solve a 26 city traveling salesman problem. Frankly, if your salesman needs to visit more cities than that, you could probably gain more by re-targeting your sales to larger markets.

Wow, if each core could...

[caywen]

If each core could model one small patch of skin on one of those busty 3D models, then, well, ... wow!

Re:How many problems can these systems really solv

[Anonymous Coward]

Sorry but throwing more processing power at crap software doesn't help... Microsoft realised that recently. :P

100 million cores not so much...

[w0mprat]

Exascale computing may seem mind bogglingly implausible at first glance, but one forgets that logic switch density goes up with the square of the process size reduction. A 1000-fold increase in computing is merely a 10x reduction in process size. Intel seems confident silicon can approach this, although it may be the realm of graphene and nanotubes.

1997/8 The first teraflops class supercomputers. We now have 32-45nm silicon.

2008/9: First petaflops class supercomputers. Today, teraflops computing is available in your desktop. A single $100 800 core GPU is theoretically a match for the 1997 #1 supercomputer.

2018/19: A single $100 ASIC should be capable of a petaflop. 3-4nm would be required to keep pace. Enter the era of exascale computing.

Oddly Moore's law detractors have been so consitently wrong, the burden of proof is now on the critic.

But...

[TheOV]

can it run Crysis???

I've often wondered...

[petrus4]

...what might happen if we could run a copy of The Sims on a truly massive supercomputer. It would need to be somewhat customised for that particular machine/environment, of course, but I think it could be interesting.

There were times when I did see something close to genuinely emergent behaviour in the Sims 2, or more specifically, emergent combinations of pre-existing routines. You need to set things up for them in a way which is somewhat out of the box, and definitely not in line with real world human architectural or aesthetic norms, but it can happen.

Makes me think; if we could run the Sims, or the bots from some currently existing FPS, parallel on a sufficiently large scale, we might eventually start seeing some very interesting results come from it, at least within the contexts of said games.

Re:I've often wondered...

[ioshhdflwuegfh]

...what might happen if we could run a copy of The Sims on a truly massive supercomputer. It would need to be somewhat customised for that particular machine/environment, of course, but I think it could be interesting.

Yes, especially because scientists don't really have anything better to run on those computers.

There were times when I did see something close to genuinely emergent behaviour in the Sims 2, or more specifically, emergent combinations of pre-existing routines. You need to set things up for them in a way which is somewhat out of the box, and definitely not in line with real world human architectural or aesthetic norms, but it can happen. [emphasize added]

I see here something close to being genuinely interesting, especially when you say that the observed behavior of the game is "not in line with real world human architectural or aesthetic norms, but it can happen." You mean: it can happen in the computer game.

Wisdom of Experience

[jd]

Average human lifespan is 80 years. Assuming the author is roughly 30, then the author need only fear being ridiculed for underestimating the future for the next 50 years. Assume Moore's Law continues to hold and assume that scheduling problems constrain individual motherboards to 16x16 cores (16-way SMP is your limit, it's hard to imagine hardware inside the CPU is going to be any easier than hardware outside the CPU).

This means a desktop system in 50 years time can realistically expect to be limited to the equivalent of 256 cores that are each running at 100 terahertz. It won't be quite that architecture, but it should have that level of power.

It seems very wise to be scoffed at by a few people now and then hailed as a Visionary in his old age. It's not like anyone would give him anything now for being accurate, but Visionaries get their own TV shows, awards, endorsements - serious cash!

Far, far better to be a Visionary.

Re:How many problems can these systems really solv

[jd]

Easy solutions =have= been found, but then Inmos was sold off. The mice were furious.

Re:How many problems can these systems really solv

[Anonymous Coward]

Nice reference.

What are these computers used for, though? Weather simulations? Building bombs?

Re:How many problems can these systems really solv

[jd]

Designing nuke reactors, designing aircraft (transsonic flows are nasty, hypersonic is a nightmare), designing bombs, weather simulations (or so they say), designing racing cars (almost all Formula 1 teams own or rent supercomputer time and Bloodhound - the 1000 MPH car - can't be designed any other way), processing data from particle accelerators (just collecting terabits of data per second isn't easy), designing new generations of microprocessors and pattern analysis in genetics.

Of these, I'm most familiar with processing data from accelerators. The work I did at Daresbury (20 years ago, back when there was still an SERC) involved collecting data from a very tiny accelerator. A mere 20 MeV. The design was over-specced by a bit (they wanted to be able to handle 2^65536 32-bit words of data) but the system really did max out both the processing and networking capabilities that existed at the time. And Eurogam was a small, small project by nuclear physics standards. The LHC, if they can ever keep it running, will be logging terabits of data per second. The European high-performance grid exists because the alternative would be convoys of 40' trucks hauling SANs. And you know how the French drive.

Each core simulates a neuron....

[jameskojiro]

Make a Beowulf cluster of these and violla, simulated brain!

What the hell...

[tengeta]

Chances are just like 10 years ago, the architecture won't even be the same. This is useless, 100 million cores are only needed by google for its porn searches.

Storage?

[mattr]

If you are fighting about 1000 vs. 1024 cores, you haven't got enough of them yet.

10^8 cores isn't that much.
Human body: 10^14 cells, 10^11 neurons, 10^14 synapses.
It would be enough to simulate a brain maybe if each core simulated 1000 neurons and it is interconnected as well as a brain. Basically if it's a brain.

You could simulate a brain at 1000 neurons per core but it has to be cheap enough, small enough, low enough power consumption and dissipation, well enough interconnected and - okay basically you have to have a brain.

It would be very useful in biology, though even at the recent petacomputer discussions there was question about whether data should really be stored, it is so expensive to do so. Ideally you would put a drop of blood in and the data would be driven in real time through the system, which would
The problem is data storage. I was in a seminar about the petacomputer being built in Japan. The people were saying that there is a real question about whether data should be stored and how.

power is constrains exaflop computing

[peter303]

If each core consumes 10 watts, which is small for an Intel system, you are really talking about a gigawatt then. Fortunately, this is a the same problem as mobile computing, and innovation is converging to solve both.

1999 called...

[mrraven]

They want their Matrix fan back...

Wait...

[DanielSmedegaardBuus]

That settles it. I'm NOT purchasing any new hardware until 2018.