Economics of Moore's Law

[WiCKeD]

The Register looks at a study by iSuppli analysts that reviews the economic viability of Moore's Law. Battling on in some form for years, it will be interesting to see whether it adapts again.

“The usable limit for semiconductor process technology will be reached when chip process geometries shrink to be smaller than 20 nanometers (nm), to 18nm nodes,” said Len Jelinek, director and chief analyst, semiconductor manufacturing, for iSuppli. “At those nodes, the industry will start getting to the point where semiconductor manufacturing tools are too expensive to depreciate with volume production, i.e., their costs will be so high, that the value of their lifetime productivity can never justify it.”

"The semiconductor industry will be living with historical generations of technology longer than it did before," Jelinek says. "You are not seeing these geometries rise and fall off the way they did before. Rather, they are living on."

And they are living on because chip makers are going to be forced by the high cost of each generation of chip technology to maximize for money generated by a process instead of chip performance and lowering the cost of chips. "Historically, the focus in the semiconductor industry was always how quickly you could move to the next geometry node. Now the question is how to make money by sustaining a specific node."

Full Article @ The Register

[S3CUR3 1]

Interesting... this was bound to happen sometime.

[untrueparado]

lol, one day processors wont be able to grow in performance without getting bigger in size or energy consumption, but we wont see that for another 20 or so years. wonder what intel is gonna do after that?

[slash_2_2000]

So basically the smaller the chip the closer to bankrupt the company goes. Doesn't overly matter though, with carbon chips on the horizon. Bye, bye silicon!

[Zer010]

Originally Posted by slash_2_2000

So basically the smaller the chip the closer to bankrupt the company goes. Doesn't overly matter though, with carbon chips on the horizon. Bye, bye silicon!

Sort of, the other point of the article is what is done with the hardware itself. Anyone that has worked with PThreads before will know the pain of parallel coding.

I did recently hear from my IT manager that Virtualization software is getting to a point to where you can dedicate entire cores to applications, that would be great honestly.

[SCOLANATOR]

Will be interesting to see what future developments bring, its has to happen at some stage. There was an article over at the inquirer showing a transistor a university had made being one atom in size.
~scol

[ELItheICEman]

Originally Posted by SCOLANATOR

Will be interesting to see what future developments bring, its has to happen at some stage. There was an article over at the inquirer showing a transistor a university had made being one atom in size.
~scol

Okay, is this even possible? Because it's not immediately obvious to me as to how they'd go about producing any such thing. AFAIK a transistor has to be made out of a semiconductor, which is a material made up of atoms.

[Maniac]

Originally Posted by ELItheICEman

Okay, is this even possible? Because it's not immediately obvious to me as to how they'd go about producing any such thing. AFAIK a transistor has to be made out of a semiconductor, which is a material made up of atoms.

well... a simplified transistor is basically just a switch... I found the article

http://www.news.cornell.edu/releases...istor.deb.html

[Deluded]

Yeah... I reckon that making chips that go faster will become increasely impractical when you make it smaller and smaller.

I think in the future we'll just get CPUs that are multi-cores, like the i7.

Or Intel/AMD would make one of those xxx core CPUs where 2 or 3 core are rated at 4 or 5 or even 6 Ghz for very complex problems that cannot be parallelized (sp?) and dedicate the remaining cores (maybe rated at 1.5-3 Ghz) for problems that can be parallelized.

Of course, this is just my idea of what would happen to the future chips, since we're already reaching the limits on how small the transistors can become.

[SSPrncVegeta]

Originally Posted by Deluded

I think in the future we'll just get CPUs that are multi-cores, like the i7.

How do you expect more cores to be added if they aren't shrinking in size?

Also, don't forget that not everything can be parallelized which limits the performance gain from more cores.

[Deluded]

Originally Posted by SSPrncVegeta

How do you expect more cores to be added if they aren't shrinking in size?

Also, don't forget that not everything can be parallelized which limits the performance gain from more cores.

This is, of course, into the future. Bear in mind that current programs that take advantage of multicores processing are in a very small minority. In fact, multicore processing is still in it's infancy (I am refering to home/business multicore processing, not supercomputer multicore processing).

I was speaking about the time in the future when more than 50% of programs are optimized for multicore processing. That's what I think will happen because more computers are sold that are Duo/Tri/Quad than before. Obviously the programmers will want to take advantage of all that processing power avaliable.

I am not forgetting that writing a program to work on multicore is very difficult, but it can be done and will eventually be mainstream IMHO.

Still, I think that maybe Intel/AMD will go this route that I mentioned:

make one of those xxx core CPUs where 2 or 3 core are rated at 4 or 5 or even 6 Ghz for very complex problems that cannot be parallelized (sp?) and dedicate the remaining cores (maybe rated at 1.5-3 Ghz) for problems that can be parallelized.

If they are not shrinking in size, then make the die size bigger.

[ELItheICEman]

Originally Posted by Maniac

well... a simplified transistor is basically just a switch... I found the article

http://www.news.cornell.edu/releases...istor.deb.html

This article shows that the transistor itself is a molecule, but the current only flows through one atom. That's not the same as saying the transistor itself is one atom. That's just not possible.

Glad to hear they're finding this stuff nearby me though. Cornell's only ~1.5 hours away. Kinda cool.

[SSPrncVegeta]

Originally Posted by Deluded

I am not forgetting that writing a program to work on multicore is very difficult, but it can be done and will eventually be mainstream IMHO.

There are also fundamental limitations of how parallelized tasks in a program can be.

[Grimm]

This is only for electronic processors.

As technology develops for optical processors, we will probably find quite a few more ways to increase speed.

Imagine a processors that uses light instead of electrons for precessing. The same core could process hundreds if not thousands of threads simultaneously by using different frequencies of light.

Then the ability to discriminate between frequency and prediction are the primary determinators of how powerful a chip is. The more able it is to discriminate the more frequencies you can simultaneously read. The ability to process all available outcomes and just pick the right one will speed things up a bit. We will be back to Moore's Law.

[Deluded]

Interesting, I don't think I've heard of a optical processor; all I know about optical hardwares relate to CD/DVD/BluRay/Fiber. A link or source to that please?

Or it could be because I'm vastly behind technological advances, but still.

[SSPrncVegeta]

Originally Posted by Grimm

Imagine a processors that uses light instead of electrons for precessing. The same core could process hundreds if not thousands of threads simultaneously by using different frequencies of light.

Light may (will) increase processing speed, but thinking that adding more cores and threads will increase performance indefinitely is based on the assumption that programs can be 100% parallelized. Not everything can be made parellelized. Depending on how much of it can't be determines the benefit of adding more CPUs/cores/threads. We're already experiencing diminishing returns with four cores on programs that are 75% parallel, with the maximum performance being four time faster than a single core. 90% starts slowing down before 16 cores, and 95% at 32 cores. Seeing as how duals came out in 2005, quads in 2007, and eight due out this year, we can expect the number of cores to double every two years. By 2013 we will have reached the maximum marginal benefit of adding cores. A light CPU with 65536 cores would only be 17% if you disabled all but 128 cores, and that's assuming you can even all programs 95% parallelized.

[Kevster2004]

Only article i could dig up on these optical processors, quite awhile ago though at 2003 http://www.newscientist.com/article/...ght-speed.html

The device is called Enlight and can perform 8000 billion arithmetic operations per second, about 1000 times faster than a standard processor. Previously this type of processor was only available to highly financed government laboratories, says Lenslet's founder, Aviram Sariel.

Sounds fast

[Grimm]

Originally Posted by SSPrncVegeta

Light may (will) increase processing speed, but thinking that adding more cores and threads will increase performance indefinitely is based on the assumption that programs can be 100% parallelized. Not everything can be made parellelized. Depending on how much of it can't be determines the benefit of adding more CPUs/cores/threads. We're already experiencing diminishing returns with four cores on programs that are 75% parallel, with the maximum performance being four time faster than a single core. 90% starts slowing down before 16 cores, and 95% at 32 cores. Seeing as how duals came out in 2005, quads in 2007, and eight due out this year, we can expect the number of cores to double every two years. By 2013 we will have reached the maximum marginal benefit of adding cores. A light CPU with 65536 cores would only be 17% if you disabled all but 128 cores, and that's assuming you can even all programs 95% parallelized.

That's only if you restrict yourself to running in parallel. But if you know process 1 will return a number between 1 and 1000 and you have to wait for the result to run process 2.... since you have the bandwidth, why not run all 1000 results and just use the one that matches the result of process 1 when it's done?

For a personal computer being able to run 64,000 processes at once wouldn't be that important. Being able to run 128 would be beneficial so that your computer could run the entertainment center, the home environmental controls, play a few computer games with the kids, compile the sports scores for dad and help mom file the taxes all at the same time.

But the server at work could make use of them...

The technology isn't as useful as it could be based on current processes, but the potential opens up dramatically when you consider that we could develop new methods to utilize that power.
Can you imagine how awesome a GPU would be that was able to make use of this?

[SSPrncVegeta]

Originally Posted by Grimm

That's only if you restrict yourself to running in parallel. But if you know process 1 will return a number between 1 and 1000 and you have to wait for the result to run process 2.... since you have the bandwidth, why not run all 1000 results and just use the one that matches the result of process 1 when it's done?

Yeah, that's when diminishing returns begin. The marginal benefit increases until specialization is maxed out (1 processor per thread). When it starts to decrease, it's because you're only benefiting from having threads being thrown on unused cores.

The point is if it takes a single core X amount of time and 5% can't be run in parallel, using more cores to speed up the 95% to take less than the time needed for the 5% is redundant. The second point is, the maximum benefit from 65535 scores is 20x. That's a number we've been experiencing according to Moore's Law (if doubling the transistor count results in double the performance). If Intel and AMD just stick with only slightly increasing individual core speed but doubling the number of cores every two years, we'll be maxing out multicore-usefulness in less than four years.

Can you imagine how awesome a GPU would be that was able to make use of this?

AMD, Intel and nVidia all realize how limited CPUs are when it comes to parallel processing, which GPUs accelerate at. That's why the goal is to merge them into a single device: a single CPU core that's extremely fast for the 5%'ers that can't be split up, and let the GPU do 95%.

[Grimm]

Originally Posted by SSPrncVegeta

Yeah, that's when diminishing returns begin. The marginal benefit increases until specialization is maxed out (1 processor per thread). When it starts to decrease, it's because you're only benefiting from having threads being thrown on unused cores.

The point is if it takes a single core X amount of time and 5% can't be run in parallel, using more cores to speed up the 95% to take less than the time needed for the 5% is redundant. The second point is, the maximum benefit from 65535 scores is 20x. That's a number we've been experiencing according to Moore's Law (if doubling the transistor count results in double the performance). If Intel and AMD just stick with only slightly increasing individual core speed but doubling the number of cores every two years, we'll be maxing out multicore-usefulness in less than four years.

You have not accounted for one of the things that interferes with most home users. Additional processes eating up too much of the CPU time. If not for the additional processes, in most cases, the current CPU power is more than enough. So by building a CPU that can run hundreds or thousands of processes at once that problem is eliminated.

Sure, for CAD drawing and rendering the system would cap out on benefit from virtual cores. But a home system could run 800 instances of Folding @ home and still pump out a killer gaming experience.

A home would have one computer and as many "terminals" as the owner wished.