Intel To Ship 48-Core Test Systems To Researchers

MojoKid writes "Just when you thought your 6-core chip was the fastest processor on the planet, Intel announces plans to ship systems equipped with an experimental 48-core CPU to a handful of lucky researchers sometime by the end of the second quarter. The 48 cores are arranged with multiple connect points in a serial mesh network to transfer data between cores. Each core also has on-chip buffers to instantly exchange data in parallel across all cores. According to Sean Koehl, technology evangelist with Intel Labs, the chip only draws between 25 and 125 watts."

Comments

I just have to ask

[toygeek]

Can you imagine a *Beowulf cluster* of these things!? Think about the possibilities!

Re:I just have to ask

[Dayofswords]

never have a loading bar again!

Re:I just have to ask

[Anonymous Coward]

Yes, I can think of the possibilities...

A Jaguar (or Roadrunner) of these processors would still be too slow to numerically solve the geomechanics problems I grapple with daily though. A Jaguar equipped with these processors would be approximately 20 petaflops peak. To simulate 1 sec of fracture of a 10mm cube of rock on the atomic scale would require of order 10^36 floating point operations. To do that would take 10^20 sec at 10 petaflops. Not bad really...that's only 10^12 years. Oh wait, the universe hasn't even been around that long...

Having said that I'm a researcher who writes and uses high-performance parallel software daily. How might I become one of Intel's select few to trial these chips? I can certainly think of ways to keep them warm!

Please Intel please! ;)

Re:I just have to ask

[MrMr]

fracture of a 10mm cube of rock on the atomic scale
Ha, I can do that in less than a second, with my serial mallet.

Re:I just have to ask

[Anonymous Coward]

take a NEC SX-9 and things gets a lot better maybe:)

Re:I just have to ask

[urusan]

It certainly beats the 10^13 years it would take with a Jaguar!

Re:I just have to ask

[Jurily]

Now 48 CPUs can wait for the disk!

Re:I just have to ask

[Cinnaman]

Talking about Jaguars and multiple processors reminds me of a certain console...

Re:I just have to ask

[WrongSizeGlass]

How might I become one of Intel's select few to trial these chips? I can certainly think of ways to keep them warm!

Please Intel please! ;)

Well, posting as AC certainly will help your chances.

Re:I just have to ask

[anarche]

Having said that I'm a researcher who writes and uses high-performance parallel software daily. How might I become one of Intel's select few to trial these chips? I can certainly think of ways to keep them warm!

ummm, lets start by not explaining why one of these things won't help your research?

Re:I just have to ask

[Anonymous Coward]

Yeah, because Intel ships goods based on /. accounts.

Re:I just have to ask

[Anonymous Coward]

Traditional disks are on the way of the CRT monitors of yesterday. Processor caches are getting huge and cheap machines came loaded with plenty of RAM.

Re:I just have to ask

[Anonymous Coward]

... too slow to numerically solve the geomechanics problems I grapple with daily though...Having said that I'm a researcher who writes and uses high-performance parallel software daily.

Ah man, how many times have we told you to stop masturbating in public?

Re:I just have to ask

[leromarinvit]

fracture of a 10mm cube of rock on the atomic scale
Ha, I can do that in less than a second, with my serial mallet.

You left out an important detail: 1 sec of fracture of a 10mm cube of rock on the atomic scale

Whoa! Chuck Norris has a Slashdot account?

Re:I just have to ask

[Anonymous Coward]

Now if the just build a motherboard that supports a 16 of these on the board I'll be happy...for a while...

Re:I just have to ask

[Jurily]

That doesn't really help me as long as people are doing disk I/O in the GUI thread.

Re:I just have to ask

[Anonymous Coward]

Here is a preprint on cond-mat intended to solve your atomistic fracture simulation problem.

http://arxiv.org/abs/1001.2456

Email the second author (she has a website at uni Luxembourg) and she will provide the a slightly more readable later revision of the paper. If you ask nicely we might even let you have beta code.

JTB

Re:I just have to ask

[jon3k]

Wow, modded insightful. Now funny -- that I would understand. Apparently no one has heard of QPI or SANs? Come on, guys!

Re:I just have to ask

[White Flame]

Now 48 CPUs can block on cache misses!

fixed. Memory is the new disk.

Re:I just have to ask

[mysidia]

Doesn't matter... Intel designed the chip for parallel software research and development

If you are not a software researcher researching parallel software programming architectures, this chip is not for you.

If your research is just running simulations, then parallel software development itself is not the subject of your research.

Re:I just have to ask

[AndGodSed]

Who cares, it's not like he is really THAT good... and no I am not afraid of saying that, he cannot be everywhere at once. Hullo... there is a knock at my door...

Re:I just have to ask

[haruchai]

Let me save you some time, trouble and hair - your problems are too difficult to solve numerically. Go find something else to do
unless you plan to live for a very, very long time.

Re:I just have to ask

[Douglas Goodall]

I can do that is less than a second with my cereal mullet.

Re:I just have to ask

[azmodean 1]

Now 48 CPUs can block on cache misses!

fixed. Memory is the new disk.

Don't fight guys! There are plenty of I/O stalls to go around!

But seriously, in addition to the theoretical limits, I think the kernel level handling of a NUMA that is *this* non-uniform would require significantly higher sophistication in the cache-handling code than is currently available. Then again, I do tend to go a bit cross-eyed when I try to follow lkml discussions of cache coherency handling, so maybe it is sophisticated enough to handle it ok already.

Larrabee

[TheKidWho]

I believe this is the remnants of Intel's failed Larrabee chipset which was supposed to compete with Nvidia and ATI.

A nice article on the story behind Larrabee and it's failure:
http://www.brightsideofnews.com/news/2009/10/12/an-inconvenient-truth-intel-larrabee-story-revealed.aspx

Re:Larrabee

[Anonymous Coward]

Indeed. Or, to be more precise, it's their "Bangalore" chip which is basically the same thing as Larrabee without the graphics-specific subunits (texture unit) and perhaps the fancy-pants cache coherency / ring bus architecture.

Re:Larrabee

[jmknsd]

No, actually It is basically a bunch of Pentium 3s with cache coherency removed for a small chunk of on chip RAM, and a message passing interface for inter core communication. It has alot of interesting features, and is more usable than the 80 core chip they came out with a few years ago.

Re:Larrabee

[Anonymous Coward]

Yeah. that's what I was thinking too. It suddenly becomes a whole lot less exciting when you consider that it's just a 48 core first generation Pentium rather than a 48 core i7.

Re:Larrabee

[Anonymous Coward]

That's not a particularly auspicious name for a chip. I'd assume that a "Bangalore" CPU would promise that it could get the work done twice as fast for half as much money due to "parallel architecture" - but you'd launch a program, only to discover that it actually took 10x as long, every instruction needed to be told *exactly* what to do, and the results were so full of errors that it took an additional non-Bangalore CPU working full time just to get things right.

Re:Larrabee

[Anonymous Coward]

From what I hear out the mouths of Intel employees (that are hands-on with products we buy and some you'll never hear about), Larrabee isn't considered officially dead. They're still working on it behind the scenes until they can pump out something everyone feels can go the distance for them.

Re:Larrabee

[jon3k]

HAHAHAHAHAHAHA i'm dying over here. apparently the indians got moderator points today.

Re:Larrabee

[Spy Hunter]

No, actually this is a separate effort entirely. This is a product of the same group which produced the "Polaris" 80-core chip, and is meant for research into communication models and memory architectures for massively parallel systems.

Larrabee is still ongoing as a separate project with a different focus. Larrabee is all about getting maximum throughput by adding a wide vector unit with a whole new instruction set to each x86 core. As far as anyone outside Intel knows, the plan is still to eventually release some Larrabee prototypes as-is (with the texture units and everything), and to develop a Larrabee 2 with the lessons learned that can actually compete directly with GeForce and Radeon in the graphics card market.

Re:Larrabee

[Anonymous Coward]

Which is what larrabee was.

Re:Larrabee

[Anonymous Coward]

Larrabee was not based on the P3 architecture.

bullshitter

[Anonymous Coward]

>>> Sean Koehl, technology evangelist
Oh... a bullshitter

Busted ...

[Macka]

Well, not very eloquently put, but it's obviously not just me who can't get to his own Slashdot page. I first noticed it about 11 hours ago, so it's been broken for some time.

And it runs Linux

[Macka]

According to the video they're running Linux on this thing with a custom kernel. No specific details on the changes they had to make to get it running yet.

Re:And it runs Linux

[Taco Cowboy]

According to the video they're running Linux on this thing with a custom kernel.

Is that custom-kerneled Linux open-sourced? o0

Re:And it runs Linux

[Dayofswords]

its just for the concept use of it, doubt they will release it

Re:And it runs Linux

[klingens]

If they distribute it to the researchers they must release it to the researchers or commit a GPL violation.

Of course the researchers don't want to demand source since then they won't get freebies like this the next time Intel does such a Santa Claus imitation of distributing presents.

There's an interesting thought: what happens if you are a beta tester who has to sign a NDA to get something which includes GPL code. What takes precedence? Your NDA, or your right to demand source to the GPL stuff and redistribute it publically?

Re:And it runs Linux

[solevita]

or your right to demand source to the GPL stuff and redistribute it publically?

But this is all just wrong. There's no requirement to make GPL code public, you only need to make it available to the people that receive the binaries. So the researchers will likely be given some source code, but nobody has to release that to the rest of us.

Re:And it runs Linux

[mdwh2]

http://www.gnu.org/licenses/gpl.html , section 10:

Each time you convey a covered work, the recipient automatically receives a license from the original licensors, to run, modify and propagate that work, subject to this License. You are not responsible for enforcing compliance by third parties with this License.

An entity transaction is a transaction transferring control of an organization, or substantially all assets of one, or subdividing an organization, or merging organizations. If propagation of a covered work results from an entity transaction, each party to that transaction who receives a copy of the work also receives whatever licenses to the work the party's predecessor in interest had or could give under the previous paragraph, plus a right to possession of the Corresponding Source of the work from the predecessor in interest, if the predecessor has it or can get it with reasonable efforts.

You may not impose any further restrictions on the exercise of the rights granted or affirmed under this License. For example, you may not impose a license fee, royalty, or other charge for exercise of rights granted under this License, and you may not initiate litigation (including a cross-claim or counterclaim in a lawsuit) alleging that any patent claim is infringed by making, using, selling, offering for sale, or importing the Program or any portion of it.

So the way I see it, either they have the right to redistribute, or Intel would be committing copyright infringement.

And it would be an obvious loophole if a contract, NDA, EULA or anything else could trump the GPL, because then people would just distribute a GPL application with an additional bit saying "Actually you have to agree not to redistribute this application".

Re:And it runs Linux

[klingens]

So the researchers will likely be given some source code, but nobody has to release that to the rest of us.

Yes, none of the researchers has to release anything, but at the same time they have the right under the GPL to do so if they wish. So the right under the GPL conflicts with the probable NDA they signed. Can you sign away the rights you have under the GPL?

Re:And it runs Linux

[mdwh2]

Yes but the researchers can choose to redistribute the binaries, and then any of the recipients also have the right to receive the source.

Re:And it runs Linux

[TheRaven64]

What takes precedence? Your NDA, or your right to demand source to the GPL stuff and redistribute it publically?

It's complicated. The GPL takes priority because the NDA is in violation of it, however the problem is that, by making you sign the NDA, Intel[1] is in violation of the GPL, but you do not have standing to sue them for it. The kernel developers could sue Intel for copyright infringement, by distributing their code without a valid license to do so. Intel could sue you for breach of the NDA. You, however, would have no recourse against Intel if you chose to distribute the code.

In fact, Intel could distribute the kernel under some terms other than the GPL, preventing you from distributing it legally. They would also be committing copyright infringement, but only the original copyright owners would have standing to sue, it wouldn't help you at all if they decided not to bother.

[1] Names of organisations in this post are hypothetical placeholders. I am not accusing Intel of GPL violation.

Re:And it runs Linux

[bhtooefr]

GPL says that if they don't have the right to redistribute for another reason, due to another license, then they don't have the right to redistribute the GPL code, IIRC.

Re:And it runs Linux

[Vellmont]

what happens if you are a beta tester who has to sign a NDA to get something which includes GPL code. What takes precedence?

The GPL doesn't allow further restrictions to be placed on it. If a company tried such a tactic they'd be in violation of the GPL and the copyright owner could sue. Remember that the GPL also protects the rights of the copyright holder.

Re:And it runs Linux

[Anonymous Coward]

Really though. This discussion seems a bit silly to me. After all, we want hardware manufacturers to support Linux. Who cares if they don't immediately release their source for a custom kernel while they're working out how to get it working in the first place. The engineers that are going to be working on the driver have code access. When they actually ship the chips and don't release code then I'd worry about the GPL. I can't help but feel that sometimes people get too caught up in the letter vs the spirit of OSS. It's not like anyone has any use for the forked kernel source if they don't have this experimental chip. And if they give it to you for engineering they pretty much have to give you some source. I'd much rather say, Thank You Intel, for working on cutting edge research using Linux, please just be sure to get the mainline kernel devs the source to your kernel a bit ahead of when the chip ships so that they have the chance to integrate the changes to avoid any cockups when you first pop in the liveCD after buying the hardware.

Re:And it runs Linux

[mysidia]

Well, legally, you have to follow them both, as long as they are both enforceable, unless the NDA specifies the software is not covered, or Intel somehow excludes it from the NDA.

That is impossible, so you can't distribute any part of the software in any way, ever. But a consequence of that is Intel would have violated the GPL by distributing open source software under an NDA.

Yes, a very clean-cut violation, for the person issuing GPL code under NDA. You are not supposed to be able to do that legally, the GPL is very explicit about that.

Unfortunately, Intel's GPL violation, in that case, would not be directly actionable by you, you lack standing, unless you wrote the open source program. Only the copyright owner of the open source software can take any legal action over copyright infringement.

Oh yeah, and since you wouuld be under a NDA, you are probably not authorized to report any of this to the author, or even the fact that you received the software.

If you do, you may expose yourself to liability. Even if the company was ultimately found to be infringing, the NDA could still be upheld.

And the damages for NDA violation probably exceed the damages for infringement.

Catch 22.

Re:And it runs Linux

[mysidia]

Well, you can be made to sign them away, but then it is a GPL violation for the person distributing the software to you, since they "made you sign away rights", they have in effect not granted you those rights. Therefore, they violate the GPL by not distributing under only the terms of the GPL.

The problem is, you won't be communicating the facts about that infringement to anyone, as doing so alone would violate the NDA.

If the author never learns of the breach, they can never take action or seek remedies for the breach, and proving damage would be a bit hard. Once the company became aware of the NDA violation, they would no doubt pursue action against the recipient of software, and any destroy any evidence if the breach were willful.

Unfortunately, the GPL is not a contract, and noone has to sign the GPL before obtaining software. It also does not contain terms guaranteeing to release any recipient of GPL software from you from all NDAs with regards to the software.

Re:And it runs Linux

[mysidia]

Well, that's my belief also. That text is from the GPL Version 3, which does not apply to the Linux kernel. GPL version 2 applies to the Linux kernel, it is less detailed and less restrictive, it says:

6. Each time you redistribute the Program (or any work based on the Program), the recipient automatically receives a license from the original licensor to copy, distribute or modify the Program subject to these terms and conditions. You may not impose any further restrictions on the recipients' exercise of the rights granted herein. You are not responsible for enforcing compliance by third parties to this License.

Re:And it runs Linux

[Anonymous Coward]

> What takes precedence? Your NDA, or your right to demand source to the GPL stuff and redistribute it publically?

They're both active at the same time.

The GPL is an obligation between two parties, not the public at-large. Intel is obligated to distribute the source code to researchers it provides compiled binaries. Neither party is obligated to release the source code to the public at-large because the public did not receive any code covered under the GPL.

Now, the NDA prevents the researchers from releasing that GPL code to the public or disclosing publicly unavailable modifications to existing code.

Thus, it may be distasteful from a community perspective, but there is no violation and no legal problems.

Re:And it runs Linux

[ejasons]

Now, the NDA prevents the researchers from releasing that GPL code to the public or disclosing publicly unavailable modifications to existing code.

And, it is this part that makes the code a violation of the GPL, which states that IBM may make no restrictions on what may done with the source code they distribute under it.

640 C (cores) should be enough for everybody

[youn]

maybe that's what bill gates meant when he said 640K should be enough... K as in Core .. it was a spelling mistake;)

Re: 640 C (cores) should be enough for everybody

[selven]

With the kinds of things Bill Gates did, I don't know if even 640' C would be enough.

Re: 640 C (cores) should be enough for everybody

[Oddscurity]

Or 640 Kelvin?

Re: 640 C (cores) should be enough for everybody

[TeknoHog]

Too bad multiprocessing did not exist back then, as Intel had yet to invent the Core.

Re: 640 C (cores) should be enough for everybody

[CrazyJim1]

6 Coors are typically enough, unless you throw a small party, then you'd want 48.

Re: 640 C (cores) should be enough for everybody

[anarche]

no no, he meant Kores

http://www.kore-usa.com/

coz he needed 640 peddle-pushing monkeys to power his plans for power dominion.

Re: 640 C (cores) should be enough for everybody

[RichiH]

640 k cores, you mean.

Re: 640 C (cores) should be enough for everybody

[noidentity]

So, Intel, only 592 more cores to go, then it'll be enough for anybody. Get to work!

Re: 640 C (cores) should be enough for everybody

[Anonymous Coward]

Naw, he was just a KDE advocate years ahead of his time.

Re: 640 C (cores) should be enough for everybody

[Anonymous Coward]

Nah - you'd still need 639 of them to run antivirus.

Can you imagine...

[EnsilZah]

...a Beowulf cluster of engineers awkwardly reading marketing information from a teleprompter?

Re:Can you imagine...

[mysidia]

How about a Beowulf cluster of US legislators awkwardly reading marketing information from a teleprompter?

Is the 48-core chip the same as the one ...

[Taco Cowboy]

... reported in this Slashdot entry - http://hardware.slashdot.org/article.pl?sid=09/12/02/215207 ??

More Cores

[lw7av]

I hope this is not a marketing ploy. I am more interested in thread management and so can't wait for the benchmark reports; if they are made public.

Tilera

[loufoque]

Might as well buy a Tilera if it's for research...
The only good thing about x86 is that it runs legacy Windows programs, but who cares about that in research?

Re:Tilera

[dbIII]

There's commercial geophysics programs that only run on Sparc and x86 - and some of them are dropping support for Sparc. The biggest x86 you can throw the software at gets the job done.
I'm sure there are similar stories in other fields.

Re:Tilera

[loufoque]

They're using clusters made of "on-the-shelf" hardware.
The 48-core Intel CPU is certainly not that.

Re:Tilera

[timeOday]

The only good thing about x86 is that it runs legacy Windows programs, but who cares about that in research?

Just because source code is available doesn't mean your problems are over. Even ironing out x86-64 (which millions of people can use) has taken years for the linux distros.

Re:Tilera

[oakgrove]

Just because source code is available... Even ironing out x86-64 (which millions of people can use) has taken years for the linux distros.

Interesting. See, I was running 64 bit distros when Vista was still called Longhorn. I'm also quite sure it was a year or so before XP x64 edition was released. And everything ran great. The only problem I had was flash. Of course, I had the source for everything except for, you guessed it, flash. Imagine that.

Re:Tilera

[mysidia]

If you watch the video, read the articles, it is for x86 parallel programming research, not general science research. You can't do that on a Tilera.

Re:Tilera

[dbIII]

No, but it's still very likely that the binaries would be able to run without modification. That's a very big deal when you have no chance of getting anywhere near the source code.

Correction

[Anonymous Coward]

"Just when you thought your 6-core chip was the fastest processor on the planet, Intel announces plans to ship systems equipped with an experimental 48-core CPU to a handful of lucky researchers sometime by the end of the second quarter.

Actually, the 8-core (Nehalem EX) and 12-core (Opteron "Magny-Cours") CPUs are already faster than your 6-core CPU. And oddly enough, this 48-core CPU is actually slower than your 6-core, 8-core, or 12-core CPUs. Intel didn't design the 48-core CPU to sell it. They did it as a research project/experiment to develop new ways of interconnecting so many processing cores. While there are technically 48 cores they are far less complex and slower performing than anything that Intel is shipping retail. If you go back a year or two you can find articles where Intel unveiled the CPU and talked about performance. This is simply an exercise in massively parallel CPU design, not an effort to make a faster CPU. That's why they are shipping them to researchers, so they can study and learn how to develop uses for such massively parallel systems.

Re:Correction

[Matrix14]

Actually, no one is making faster cores any more. We've hit a large technological roadblock in that area, due largely to heat dissipation issues. The fastest commercially available x86 chips have been a little under 4 GHz for about five years now. Current chip design focuses on heat and power issues and increasing the number of cores on the chip.

Re:Correction

[mdwh2]

The fastest commercially available x86 chips have been a little under 4 GHz for about five years now.

Megahertz Myth. As far as I can tell, over the last 5 years individual cores have still been getting faster, just not with higher clock speeds.

Re:Correction

[tius]

And it won't be long before the "number of cores == speed" myth starts to show it's ugly head. Communications overhead eventually has a major impact on how much computing a processor/core can achieve; i.e. memory, I/O, inter-core. In more classical parallel implementations this is strongly felt around 8 processors/cores. However, today we have the benefit of shrinking high speed packet networks to the system and chip level, so I'm not sure where the knee in the performance curve is. If we're lucky the MHz myth/limit will help linearize performance/# of cores.

Re:Correction

[Hells Ranger]

Megahertz Myth. As far as I can tell, over the last 5 years individual cores have still been getting faster, just not with higher clock speeds.

They are going faster by shortening the pipeline so you get a shorter execution time if you have branch. Also there are more execution unit. At first there was only 1 alu doing everything. Now in a standard cpu you find multiple dedicated unit for integer, logical and float operation. It allow to execute many operation in parallel, as long as there's no dependency between the data.

Re:Correction

[TheRaven64]

Intel didn't design the 48-core CPU to sell it

Actually, they did. Unfortunately, it was delayed and didn't work as well as they'd hoped and it would have been a complete flop in its original target market so they're shipping it as a research toy to try to recoup some of their investment.

Re:Correction

[Spatial]

Correct.

Clock frequency is worthless as a measure of CPU performance. Cores have never stopped getting faster.

For example: Each individual core in a 2.66Ghz i5-750 is more than twice as fast as a 3.8Ghz P4. Often many times faster than that, depending on the workload.

Re:Correction

[Down_in_the_Park]

Wouldn't they get even shorter interconnects between the cores, if the finally go into the third dimension? "Cube" would have a totally new meaning. they just have to solve these heat problems...

Re:Correction

[symbolset]

More likely they've got this widget sitting around with all the requisite engineers raring to go. But it's a wrench that fits no bolt - they need research scientists with the type of problems that this solves to put a load on it, define the scope of its use and put it to work so they can refine the toolchain and broaden the scope.

Re:Correction

[jon3k]

"Clock frequency is worthless as a measure of CPU performance."

Well you mean when comparing chips of different architectures, yes.

Re:Correction

[Anonymous Coward]

It certainly is not not a factor, if you think otherwise you're foolish. It is a factor, just like the instruction set and the architecture of the chip itself. A chip with 3.8GHz clock vs a chip with 2.66GHz in the same family, absolutely does matter, and is useful in that context. Quit making blanket statements that newbs will just repeat to their friends.

Re:Correction

[Macman408]

I doubt recouping their investment has anything to do with it; they're not going to be able to sell it to many people, and probably not for a very high price either. They also may have to worry about support costs; they'd probably have been better off canceling it a week before shipping and saving that week's worth of engineering time.

However, this little experiment (or parts of it, anyway) may end up in future generations of Intel CPUs. They want to get it into researchers' hands now so that they can find out what's good about it, and what can be improved. Researchers can develop new algorithms to split up the work from a program so that it can be run on these many cores. Without significant advances in that area, there are many workloads (including most of the ones in the consumer space) that will never benefit from anything other than a small number of screaming fast cores. With a little help from researchers now, they can have a much better product a few years down the road.

I think in the long run, a high-performance general purpose computer will have to have two different processors; one has a small number of cores, but can do out-of-order superscalar execution, has tons of cache, a great branch predictor, high clock speed, etc. All the things that make single-thread execution fly. The other has a large (VERY large) number of very simple cores that aren't particularly fast for any one thread, but can run so many of them in parallel (or quickly switching between them) that they can finish far more work in the same amount of time. If there's a big pile of work to be done, they can handle it; perfect for video, for example.

Wait a minute - most of us already have that. The fast, complex processor is obvious. The parallel one is your GPU. The golden age of the GPU is just beginning; as more developers figure out good ways to take advantage of it, we should see some very interesting advances in both hardware and software.

Re:Correction

[RAMMS EIN]

``This is simply an exercise in massively parallel CPU design, not an effort to make a faster CPU. That's why they are shipping them to researchers, so they can study and learn how to develop uses for such massively parallel systems.''

Perhaps it would be interesting to mention Azul Systems at this point. They sell systems with 108 to 864 cores, so they may know a thing or two about "massively parallel".

Re:Correction

[kanuac]

Sure.
Overclockers just like to fool around.

Re:Correction

[Spatial]

Given that parent was talking about all x86 CPUs made in the last five years, I thought that was implicit.

"sytems"

[Anonymous Coward]

There is a typo in the headline.

48 stack traces . . . ?

[PolygamousRanchKid]

So when analyzing a kernel dump caused by a deadlock with spin locks, I get to look at 48 stack traces, to find out who got what where, and who wants what they will never get?

Sounds like fun.

Ok, creative use of LPARs/Virtualization technologies could reduce the headaches. A friend of mine owned an ancient 6-cylinder Jaguar that spent more time in the repair shop than on the road. He was looking a at 12-cylinder also in the shop, when the chief mechanic commented, "You don't want that. A 12-cylinder just means 6 more headaches.

Sometimes I think about that comment with more CPUs/cores, when I am doing kernel development.

Re:48 stack traces . . . ?

[Anonymous Coward]

Only 48? Don't forget hyperthreading.

Re:48 stack traces . . . ?

[Anonymous Coward]

And precisely how do you implement spinlocks if the cache coherency is stripped off?

Video reminds me of ...

[LoudMusic]

3DFX, so powerful it's kind of ridiculous.

http://www.youtube.com/watch?v=DmaYH1F6kho

http://www.youtube.com/watch?v=ldiYYJNnQUk

http://www.youtube.com/watch?v=o72T8qQr7GE

Great ad campaign.

Re:Video reminds me of ...

[Rockoon]

3DFX did deliver great GPU's to the market. When I finally upgraded from the Voodoo3 to the GeForce 4 ti (several generations beyond the voodoo3), that Voodoo *still* kicked its ass in multi-texturing fill rate.

Make what you want of that, but I say that 3DFX went under because of mismanagement, and not because the company couldn't deliver a great product. The company did in fact deliver a great product, right up to the end.

Re:Video reminds me of ...

[Anonymous Coward]

The end, apparently, being the Voodoo 3, since the 4 and 5 were terrible cards in general.

Re:Video reminds me of ...

[Hamsterdan]

Guess what I found the other day? my 12MB Voodoo2 cards. It's missing the SLI cable, and would propably be destroyed by a modern graphics card anyways (I think my old Radeon 7000 would blow them out of the water, so a 50$ card probably will too)

I remember playing Quake 2 at ****1024X768**** on a K6-2 (266@337). I wonder if someones still hacks drivers for those cards?

Re:Video reminds me of ...

[Rockoon]

..and by terrible you mean simply not #1?

The performance differences between VooDoo's and GeForce's, even at the end, wasn't that great overall. VooDoo's were great products all the way through to the end, when they went bankrupt through some bad decision making (cutting off the OEM's is ultimately what destroyed their revenue stream)

Ghost of the Transputer

[Mr Pippin]

Sounds quite a bit like the INMOS Transputer http://en.wikipedia.org/wiki/Transputer
Wonder what version of Occam (the programming language) will ship with it?

Re:Ghost of the Transputer

[WindBourne]

All of the ones for Linux.

Resurrect BeOS

[Gothmolly]

Thread, thread, thread, thread! (Images of Ballmer hopping around). BeOS rocked back in the day because it APPEARED to be faster, because of its pervasive multithreading. Nowadays people are impressed when MS multithreads Win7 to make it more responsive. Imagine what you can do with 48 cores.

Chip trial

[Anonymous Coward]

Hi,

I'm an engineer at Intel and we are looking for a few more candidates to test our 48-core chips. Your scientific computing project sounds like a perfect fit for our trial. Please contact me (see my account info for my email address) and we'll get you in the program.

Cheers!

Re:Chip trial

[Anonymous Coward]

I'm an Intel engineer and I can't get hold of one these things.

Re:Chip trial

[Anonymous Coward]

You are not based in Nigeria

hypervisor

[pikine]

I believe a system like that is supposed to be supported by a hypervisor, which will run just one operating system per core, e.g. Barrelfish.

AMD already has this...

[Skaven04]

AMD's new 12-core "Magny-Cours" Opteron parts will be available in 4P configurations with 48 cores and up to 512GB RAM, so...::yawn::

Re:AMD already has this...

[Skaven04]

Actually 768GB RAM...12 dimms per socket (if an OEM chooses to max out the config) with 16GB DDR3 dimms == 768GB.

Re:AMD already has this...

[supssa]

Reading is tough eh. This is 48-cores per socket. ::so smug::

Multiprocessing

[krischik]

Sure about that? Sure that no Mainframe, VAX, Supercomputer had Multiprocessing at the time?

Re:Multiprocessing

[TeknoHog]

*whoosh* I wrote my first SMP code in 2001, and it was the typical thing to do in scientific computing, had been for decades. Thus I occasionally like to comment on the recent years' "multicore" marketing phenomenon, where even some developers seem to think they have a completely new problem and they need completely new tools.

Re:Multiprocessing

[Anonymous Coward]

Mainframe, VAX, Supercomputer had Multiprocessing at the time?

No

Intel actually developed the first multi-core CPU and multi-processor systems at the behest of Steve Jobs as a condition for migrating OSX to the x86 platform. Further, it is speculated on good authority that Jobs personally headed up a crack engineering team sent to Intel expressly for the purpose of transitioning their fabs from the netburst to the core architecture. Seriously, study and learn.

Posted anonymously from my iPad at the Starbucks in Cupertino. You know the one.

Re:Multiprocessing

[krischik]

So your comment was meant sarcastic. Well yes then it all makes sense. It was just that the joke was lost on me.

Re:Multiprocessing

[jon3k]

SMP doesn't necessarily mean multi-core. And I'd call 6 (Intel) and 12 (AMD) core CPU's officially a "multicore phenomenon".

Re:Multiprocessing

[Anonymous Coward]

*whoosh*: Twat. We no longer do that.

1985, if we're bragging.

And I'm absolutely sure SMP could be dated to well before that.

- not usually AC, but parent deserves same.

--

Re:Multiprocessing

[Anonymous Coward]

SMP has been around a long time. I remember in 1993 and 1994 that people were trying to do their best to get code working on 4 CPU SGI Indigos.

Re:Multiprocessing

[Anonymous Coward]

getting code optimized for cores on the same die as opposed to smp is significantly different. Although it isn't new or unique, just more widespread to consumers now. So you can get off your "i've done smp programming" high horse now.

Re:Multiprocessing

[TeknoHog]

SMP doesn't necessarily mean multi-core.

That was kind of my point, as I had a dual P3 system way before this "multicore" stuff, and I knew it was not particularly new or special even then.

Re:Multiprocessing

[jon3k]

I think there was some sarcasm I missed? I'm thinking that was the point of the proper name "Core". It's starting to sink in now ...

Supercomputer Only? I think not!

[NonSenseAgency]

Paul Lily's assertion: "That doesn't mean that you just wasted $1,100 and that your Core i7 980X is suddenly obsolete. As part of a research project, the 48-core part might never become commercially available, and if it did, it would be destined for mainframes and supercomputing tasks, not home desktops." is just plain wrong. Anyone who has been around computer development for any time at all knows that today's supercomputer is tomorrow's desktop. While it may not be this exact CPU, sooner or later 48 core CPUs or 96 core, or....WILL be on desktops.

48 cores

[nurb432]

And still one external memory bus.

Re:48 cores

[Anonymous Coward]

Actually, 4 DDR3 memory controllers, each of which can independently talk to a bank of DDR3 memory.

Re:48 cores

[randyleepublic]

Booth was a dupe of the Fractional Reserve Banking conspriacy. So was Oswald.

Re:48 cores

[supssa]

Why is this incorrect comment at +5? Basic reading of the article / understanding how DDR3 works before modding please?

Improper Adverb Usage

[DesertNomad]

Alas, there's no such thing as "instantly", especially in multi-processor core systems. It takes all too long to move data around.

Just get a Sun/Oracle T2+

[Fallen Kell]

Seriously... That has 64 cores in one CPU. Sun/Oracle sells several systems which use 2 or 4 of these in a single system.

48 Core Mac Pro Rumors

[Macrat]

Time to get the rumor mill going....

asynchronous processors

[Anonymous Coward]

what happened to asynchronous processors? still waiting

Intel was not first to market the mesh cpu

[Anonymous Coward]

Intel did not originally design this cpu.

The 64-core mesh cpu has been sold since 2003 from http://www.tilera.com/products/TILE64.php

Here is a video of one of the founders who originally designed the mesh cpu.
http://mitworld.mit.edu/video/671

They have a 120 core cpu coming out soon.

Point is, Intel has found a way around their patents to design their own mesh cpu.
It kind of sickened me to watch Intel take all credit for this mesh design.

I can provide FIVE MILLION 48-CORE chips

[cpotoso]

I am the widow of NGALA MTUR NGILI, ex-director of Intel-Nigeria. When he left intel he managed to stash a FIVE MILLION 48-CORE CHIPS in a box in Switzerland. If you help me get these, you may keep TWO MILLION OF THOSE CHIPS FOR YOURSELF. NIGLIA MTUR NGILI niglia@hotmail.com

Asian racism, the norm at slashdot

[Anonymous Coward]

Yep at the holy slashdot there are double standards for 'evil'. Spreading FUD about our asian competition helps us keep our jobs and makes us feel better about ourselves.

Having work professionally with Indian IT workers, I have found the opposite of your intended message to be true and *in MY experience* I have found them to be infinitely more interesting than their American counterparts who can only see the world in swift monotone judgements.

Re:Asian racism, the norm at slashdot

[Anonymous Coward]

Having work professionally with Indian IT workers, I have found the opposite of your intended message to be true and *in MY experience* I have found them to be infinitely more interesting than their American counterparts who can only see the world in swift monotone judgements.

Are you talking about Indians who are working here in America, or Indians in India? In my experience, I've found the former group quite capable indeed, while the latter a complete waste of time. I suspect that the big companies like IBM end up hiring the best engineers in India, leaving those not very well-qualified for the other outsourcing duties.

Finally

[the grace of R'hllor]

Finally a CPU that might let Battlefield: Bad Company 2 run at an acceptable framerate.