Economics of Moore's Law

[SSPrncVegeta]

I've been referring, as does the Wiki, that the graph only applies to a single program. If you have a bunch of 95% programs, it scales, but even then the maximum needed is 128 per program. However, even then, the performance difference between 256 and 128 for two 95% would be from 17x a single core to 19x, ~12%.

[levathar]

In my opinion, the increase of cores by processor is good because many things can be done with the extra cores. Just like GPU do these days to folding and science research, not all science can be researched in GPU (many BOINC projects use just the CPU cores).
I believe the server market also apreciate the increase in core count.

Despite it may be a slowdown in the next coming years( we already see it maybe not jut caused by economics) I am sure that there will be research for novel tech that will surpass these problems.

[SSPrncVegeta]

Think about it though. The GPU is vastly superior to CPUs in regard to code that can be run in parallel, and the gap is increasing. The one major benefit is running sequential code. Read over this and tell me what you think.

[Deluded]

Originally Posted by Deluded

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.

But I think the multi-core trend on CPUs will take a different turn. Instead of increasing number of standard complex cores, you'll probably see just a few standard cores optimized for highest performance of sequential code. They'll execute all the hard to parallelize code. Then you'll have a large array of simple CPU cores, Larrabee style, which are individually not as fast but you can instead have many of them, which will take on CPU tasks that are mostly parallel. Including both simple and complex cores, rather than doing one or the other, is motivated by Amdahl's law. Everything can't be parallelized, so let's keep at least a few complex cores that can take on those tasks instead of constraining the whole system's scaling to the performance of the simple cores. Then finally you'll have the GPU where you do specialized super-parallel work, in particular (of course) graphics.

This is what I was trying to say.

[SSPrncVegeta]

To which I replied...

Originally Posted by SSPrncVegeta

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

The article is saying to use a handful of fast CPU cores, and let the GPU do all the parallel work. You're saying to have some fast CPU cores and lots of slower CPU cores. The slower cores would be like trying to improve performance by adding DDR1 to a DDR3 and think "more is better." It'll only bring up price and slow down the fast CPU core's potential. The amount of fast, complex cores could be 16 for all we know, but anything more would be handled quicker by a GPU of the same era.

edit- Added last sentence.

[Deluded]

****it, lol. I'm not properly informed enough in the Amdahl's law to make a solid argument like that.

Either way, I imagine that if a GPU+CPU combo would be coming out in the future, we could expect to cough in excess of $1000+ just for that CGPU.

285 combined with i7, anyone?

[SSPrncVegeta]

The speedup of a program using multiple processors in parallel computing is limited by the sequential fraction of the program.



There's a limit to how fast the overall program can perform depending on how much of it requires sequential processing. I believe the best example of two extremes would be games and database queries. The optimal combination is to let the GPU do the parallel code and the CPU do the sequential code. There's a benefit to quadcore CPUs because developers are still making the CPU do parallel code, but it wouldn't benefit nearly as much as having the GPU do it. Physics, AI, raytracing are all moving to the GPU. nVidia is prepared, Intel is preparing, Idk what ATi is doing with their Stream technology.

[MrObvious]

GPU+CPU combos are where the market is going anyway.

[levathar]

I do hope they can go up to 16 or 32 cores per chip and sell it at today's quads atleast in 10 years maybe?

[zollen]

I do hope they can go up to 15Gz single core CPU and sell it at today's quads.

Should innovation be considered in this discussion?

[SSPrncVegeta]

Originally Posted by levathar

I do hope they can go up to 16 or 32 cores per chip and sell it at today's quads atleast in 10 years maybe?

Yeah, just ignore everything I've said and 32 cores sounds great!

Originally Posted by zollen

Should innovation be considered in this discussion?

Innovation is the discussion.

[zollen]

Have you heard of any future potential technologies (i.e. nano, DNA, multi-states transistors..etc) that could bypass Moore Law? I have been looking for articles about these tech.

[SSPrncVegeta]

They'll have to invest in nano, DNA, multistate transistors to keep up with Moore's Law. They may exceed Moore's Law, but current tech can't keep getting shrunk.

[Zer010]

Originally Posted by levathar

I do hope they can go up to 16 or 32 cores per chip and sell it at today's quads atleast in 10 years maybe?

The more logical question that comes to mind is how long will it take developers to catch up to the idea of parallel programming.

I just finished this article:

Originally Posted by SSPrncVegeta

Think about it though. The GPU is vastly superior to CPUs in regard to code that can be run in parallel, and the gap is increasing. The one major benefit is running sequential code. Read over this and tell me what you think.

When I worked on the Mutex Library and PThreads I had a horrible time dealing with different code paths and optimizations however like the article stated above once I got it down I can quickly take it to another library (that doesn't mean the implementation will be correct nor the same)

Right now I am very interested in seeing how OSes will start using multiple cores with multiple threads especially with virtualization (that segs into program boxing too) in a more serious manner. Frankly they will set the bar on where things will go imo (MS with their money and OSC with their will power).

[zollen]

Originally Posted by SSPrncVegeta

They'll have to invest in nano, DNA, multistate transistors to keep up with Moore's Law. They may exceed Moore's Law, but current tech can't keep getting shrunk.

I think it is understood that current tech and architectures are heading to an technological dead end. I am sure big names like Intel and AMD have already investigated alternatives for many many years. They understood the problem better than us.

[Fr0sty]

what about a bulldozer like cpu... the rumors explained some special instruction set that could help single threaded app being procesed like they were multi threaded app and the next step would be some more complex cpu building material to pass electricity faster etc...



after reading the article that ssprncvegeta pointed it sure looks like the best setup in the futur would be a cpu + powerfull gpu + cpu like switch ..... the switch would be able to direct the right thread of work to be processed by the gpu because of their powerfull nature etc.. and the more cpu-like thread to be done by the cpu and a special motherboard could be made for the server or workstation market... like an entry level 2p workstation or 4p to 8p so now the botleneck is bandwith pci-e 4.0 anyone???? maybe such a workstation would be available by that time

[Grimm]

Originally Posted by SSPrncVegeta

Yeah, just ignore everything I've said and 32 cores sounds great!

Innovation is the discussion.

When running 8 processes that are an average of 75% parallel code, it would be awfully nice to have 32 cores.

Right now I am running 12 processes, this is normal.

[SSPrncVegeta]

Originally Posted by Grimm

When running 8 processes that are an average of 75% parallel code, it would be awfully nice to have 32 cores.

You're fighting over 2 to 4 times the CPU performance of a single core. The performance which doubled roughly every two years before multicore CPUs came out.

[Catalina588]

First, we all know that Intel is delivering 32 nm chips this year. They alreay have 22 nm fabrication underway for prototyping the next generation. So, we'll definitely get to ~20 nm in the next couple of years. Therefore, Moore's Law is good for at least the medium term.

Second, iSuppli ignores manufacturing process. Look at what hafnium and high-K did to the Penryn generation of Intel chips, which in turn allowed a better architecture, Nehalem, to generate a much higher clock. There are hundreds of thousands of Core i7 920's running air cooling at 4 GHz today. I would expect IBM and Intel to continue to push the facbication envelope.

The cost of each new generation is indeed escalating. The likely outcome is a further decrease in the world's supply of leading-edge foundries who can afford to keep on the Moore's Law curve. But that trend has been around for at least a decade. AMD, as one example, got off the manufacturing train and became a soft fab company. That move alone will not kill off AMD, in fact, it might allow AMD to survive.

PSK

[zollen]

Some really smart people in white coats have discovered a new material that is harder than diamonds and conducts electricity 100 times better than silicon.
Sounds great right? Did I mention it is also only one atom thick?
Once you put those together you have the makings of something to set the computer and IC world on its ear.
The new material is being called graphene and is pure carbon. It is being earmarked for use in touch screens, solar cells and of course high performance computer chips.
The time when we will have graphene in our computers is still a long way off though. As it stands right now scientists have only recently figured out how to make it efficiently. Before they actually had to mount carbon atoms onto tape and then peel the take away to get the graphene layer.
Still there is hope that in the near-term hybrid graphene silicon devices will popup and improve the current silicon only products we use.

Update*
The prinicple behind Graphene was first theorized in 1947. 2004 was the first time someone actually pulled a 2D sheet of graphene from a graphite composite. This was proven conclusively to be possible in 2005.
In 2007 the first transitor made of Graphene was produced
In 2008 a new method for creating Graphene was found.



http://www.tweaktown.com/news/12759/...ond/index.html