Intel Recommends Developers Plan for Massive Multi-Core Processing

[CWallace]

With so many cores, it would be interesting if future compilers use the cores themselves to determine things like pipelining, out-of-order execution, and such that are now handled by specialized hardware in each core, resulting in more complex systems. On a four-core system, for example, one core could be dedicated to ensuring that the other three cores are properly "fed" to maximize their throughput, resulting in less complex cores, overall. This would lower transistor count, which would reduce production cost and waste heat, allowing for cheaper cooling systems and more compact systems.

 

[k2spitfire88]

At the very least, it will be interesting to see how this plays out. I hope that Snow Leopard is able to do this whole multi-core processing thing right, and if it is, it will be simply amazing.

 

[longofest]

There is actually a sorting algorithm (merge-sort) that does just that - it breaks the list to be sorted in to 2 chunks that are then broken into 2 chunks till the chunks are only 2 values big, it then sorts them and puts the chunks back in order

As a result this kind of algorithm will lend itself very nicely to this kind of problem and multiple cores full stop.

The real problem with having a huge amount of cores is that it becomes nigh on impossible to program for it because as stated before in the thread some tasks just can't be broken down into tiny bits...

I know. I was describing parallel merge sort. My point is that the overhead associated with spawning so many threads for such little work to be done will be more than any gains that you will have by parallelism.

If you're a programmer, try it using simple program using pthreads (if you have a multi-core mac). If it's a dual-core, use a list of only 4 elements... try sorting with a simple bubble-sort and then use the parallel-merge that uses a bubble inner-sort.

I already did it in school, so I already know what the answer is

 

[Amdahl]

I hope that programmers begin to do this. Is there a theoretical limit on how many cores there can be?

There is Amdahl's Law, which puts a limit on how many you can use effectively.

Hardware designers will also experience difficulties creating memory switch architectures that can handle 16+ cores accessing the same memory.

 

[twoodcc]

wow, now this is some serious stuff. looking forward to seeing tons of cores!

 

[dwsolberg]

New things

I know. I was describing parallel merge sort. My point is that the overhead associated with spawning so many threads for such little work to be done will be more than any gains that you will have by parallelism. If you're a programmer, try it using simple program using pthreads (if you have a multi-core mac). If it's a dual-core, use a list of only 4 elements... try sorting with a simple bubble-sort and then use the parallel-merge that uses a bubble inner-sort.

It's clear that using old (current) methods will not work with the new multiple-core technology coming up. It's up to people to figure out clever ways of adapting. Whenever I think of these types of problems, I always think of various puzzles that seem impossible until you see the solution. My hope is that Apple is helping developers by using their brainpower to think up some solutions.

 

[TwinCities Dan]

This is exciting news!

(the times they are a changin')

 

[jons]

I have been waiting for this for years. Now if we could just get HDD transfer rates up there...

 

[dwsolberg]

Amdahl's Law

I appreciate the Wiki entry on Amdahl's law, but the page has a serious flaw: It assumes that the smallest possible chunk the task can be divided into is 1/20 (95% of a task). Naturally, that corresponds to a (almost) maximum speed increase of 20x. If a task could be divided into 100 chunks, then the maximum speedup would be 100x using the formula.

Here's a list of different sized chunks along with the corresponding speedup using 512 processors:
2000 chunks = 407x speedup
1000 chunks = 338x speedup
500 chunks = 144x speedup
100 chunks = 83x speedup
50 chunks = 45x speedup
20 chunks = 19x speedup

Clearly, the big challenge is how to solve the problem of making the chunks smaller. It seems impossible, but I doubt that it is.

 

[bug67]

"...start thinking about tens, hundreds, and thousands of cores..."

Holy smokes!

 

[hayduke]

http://en.wikipedia.org/wiki/Amdahl's_law

The interesting thing about that link is that it basically says that for code that is 95% parallelized you would need 2048 (perhaps even 4096!) cores to get a 20x speed improvement. Now, I'm not saying 20x is no big thing, but that is a LOT of cores...sounds like this is the direction we are headed though...

http://fastra.ua.ac.be/en/index.html

 

[jackfrost123]

Are we supposed to be excited over this intel? About your inability to bring something decent to the table? Anyone who's done some basic computer science will sneer at these news. Make better chips not more cores...

 

[AppleMojo]

Are we supposed to be excited over this intel? About your inability to bring something decent to the table? Anyone who's done some basic computer science will sneer at these news. Make better chips not more cores...

Just read my tag line.

 

[Amdahl]

I appreciate the Wiki entry on Amdahl's law, but the page has a serious flaw: It assumes that the smallest possible chunk the task can be divided into is 1/20 (95% of a task). ...
Clearly, the big challenge is how to solve the problem of making the chunks smaller. It seems impossible, but I doubt that it is.

No, you're just misunderestimating the problem, because the Wikipedia graphic is using an exponentially compressed X axis, but a standard Y axis. The problem is the 5% part that can't be parallelized, not how many chunks you slice the 95% that can be parallelized. If 5% can't be parallelized, then a 20x speedup is the limit of performance, with tremendous number of cores needed to get that last 1x of speedup. Practically speaking, you would stop at 64 cores for a 15.5x speedup.

This graphic demonstrates a problem where 90% is paralellizable, and 10% is serial. Maximum speedup is about 6x, with seriously diminishing returns at 8 cores (~4.75x speedup), and 9.17x speedup if you go to 100 cores: 'Linear' is perfectly parallelized.

 

[tenks]

Are we supposed to be excited over this intel? About your inability to bring something decent to the table? Anyone who's done some basic computer science will sneer at these news. Make better chips not more cores...

Right because Intel is the ONLY company in the world that makes x86 based CPUs, and apple didn't choose them because they had the best product and roadmap out of all the rest...

Darn you intel! Stop making really great chips with a great price/performance ratio! Stop developing the best desktop CPUs in the world! Darn it! I mean, cause Core 2 duo's and quads are really slow and saturated. Make better chips (oh wait they make better chips then their competition, AMD) and not more cores! Because we all know its about single cores and GHZ! Darn it Intel, we don't need quad cores that run at 4ghz on air easily, we need 1 core a 10ghz!!!


/sarcasm

 

[jackfrost123]

@applemojo Yeah champ you got it...whatever....

@trenk This is just trash talk. Obviously you don't understand what I said. Good try though. FIY at the moment they make equally good or just marginally better chips than amd, nothing more. But somehow I have to program a simple algorithm that runs as fast on one core as on 10 billion ones and get excited about it. Of course you clown 1 core at 10ghz is better than ten, but they can't do it...

btw, @ both of you bozos, see the f. graph that user amdahl posted and try to "ged it"

 

[jmadlena]

Are we supposed to be excited over this intel? About your inability to bring something decent to the table? Anyone who's done some basic computer science will sneer at these news. Make better chips not more cores...

We've kinda hit the GHz wall, or what have you. You can only fit so many transistors on a chip, why not double the cores?

Anyone who's done some basic computer science will sneer at these news. Make better chips not more cores...

I'm pretty sure the experts at Intel know a few things about computer science... more than you or I.

 

[Amdahl]

I'm pretty sure the experts at Intel know a few things about computer science... more than you or I.

Indeed, they do. But that doesn't change the fact that they need to sell chips, and the only chips they can make are the kind that need parallel code to be useful. So they are going to sell them until people don't want them, or a new use for them is discovered. Today, they are useless. Tomorrow, we'll see.

The success of these chips is not going to be in making our current computer experience faster; the success is going to come in finding ways to soak up tremendous amounts of CPU in new ways: speech recognition; visual recognition; other data pattern recognition; video encoding; 3D displays; etc.

 

[jackfrost123]

Naturally if you hit the proverbial ghz wall you add more cores, not much more to do.

I am a computer science major. I know they know more than I do, but I know equally well when they do marketing speak because they've hit a dead end and they are trying to portray it as some sort of triumph. It's of course much better than a year or two or three ago, but that's technology. That's what I am saying.

//very well put by fellow poster amdahl

 

[iceman1234]

-unless the program you are using is optimized for duo, quad, "octagonal" or... cores, you processing speeds will not drastically improve with more cores.
-this is half the reason why windows runs like a dog; windows is, and will never be optimized for each and every computer configuration.

P.S. :
-Will the universe collapse when computers have hexadecimal (16) processors and run in the hexadecimal system... Its a paradox I tell You!!!!
-That guy's graph (above) looks like it was made in Word '97 on a PC!!!

 

[kaiwai]

Wow...16 cores?? Technology is advancing so fast!

*creeky old man voice* I rememer not too long ago I used to boast about having a 7.14Mhz 68K processor in my Amiga. Oh the good old days

 

[WillMak]

So does this mean my Core "1" Duo on my Rev.A macbook still has a chance to be faster with the right software? Or are the core duos just so obsolete now?

 

[jackfrost123]

no it can't be faster. But this is to do with core 1 and core 2 being different architectures, rather a progression from 1 to 2. The duo has to do with the no. of cores: duo, quad, etc.

 

[MacFly123]

"the basic idea…program for as many cores as possible, even if it is more cores than are currently in shipping products. "

Apple has this in mind. It looks like Grand Central, and also OpenCL should helpfully take some of the weight off developers' shoulders regarding optimally using multiple cores and multiple processors...

The great thing is, that you can buy multi-core systems now, and then get a performance boost when Snow Leopard comes out next year.
Nehalem chips due soon will surely be being set up by Apple to be in line with getting a decent performance boost from Snow Leopard. Improving performance on a Core 2 Duo is easy. A dual socket Nehalem with 8 cores each? That's what Snow Leopard seems to be aiming squarely at (and it doesn't stop there - seeing as it's more about multiple cores, than using more than 2 sockets currently).
In terms of theoretical limits - Intel already has protypes with 100s of cores on.

Can't wait to see what kind of performance improvements Snow Leopard brings. Even on older machines (with mutli-core procs) there should be significant gains, I should think.

The interesting thing is - multi-cores isn't potentially the main course. If Apple can effectively crack using GPUs for General Purpose computing, then there will be an insane speed bump coming our way (primarily for desktops i'd imagine due to the thermal issues).

And if you want to extrapolate, imagine what it could do in a couple of XServes, or using Snow Leopard Macs linked up into a loose distributed processing system.
https://forums.macrumors.com/threads/511801/

I have to say I am very excited for Snow Leopard and I am also very proud of Apple for laying the foundation for a great, stable, and fast future. I still think they will sneak in at least a few cool features but nothing near the 300+ in Leopard. Either way I am very excited about what will be under the hood.

Does anyone know what Microsoft is doing in Windows 7 concerning all this under the hood multi-core and GPU processing etc???

It is looking like Apple is going to be looking even better to the public when a weird and buggy repackaged Vista called 7 is going up against the solid, lightning fast, user friendly, beautiful Snow Leopard. Microsoft Strike 2 haha!

 

[Rhalliwell1]

Buy now?

I was thinking about buying a Mac Pro within the next month or so. At the moment i really don't need the 8 core power but i am just future proofing. I am now worried that if intel start releasing processors with 128 cores or something stupid and developers start producing apps that make use of 128 cores then my 8 cores are going to be worthless and i am gonna need a new computer in the near future (not the plan! lol)

What do you all think? How far away are these HUGE processors?

And regarding the 16 core mac pro, will that be 16 core upgrade or 16 core standard?

Thanks.