Remapping Computer Circuitry to Avert Impending Bottlenecks

[Simfish]

http://www.nytimes.com/2011/03/01/science/01compute.html?_r=1&pagewanted=print

So - a question. Once this infrastructure is in place, do we have to revise all the algorithms we use? Just as we have to revise all of our algorithms when we're programming for parallel processors?

 

[caffenta]

I don't see any major change described in this article. Just a lot of buzzwords.

One thing is completely wrong in this article: multiple cores were not added because more real-estate became available. That is completely ridiculous. You can make the die as large as you want. Multiple cores came into existence because of excessive heat generation: the clock speed could no longer be increased, so extra cores were added to increase processing power.

And the absence of a “major breakthrough” is simply because the current technology was/is still good enough. “Computer designers” will not make the change. Semiconductor manufacturers will. That's only going to happen when it makes sense financially. Fabs are really expensive as is.

EDIT: there were major breakthroughs by the way. Copper was a huge one around the turn of the millennium. High-k/low-k dielectrics also. So Mr. "Computer Designer" should probably not talk about things he has no clue about.

Regarding the algorithms, the math is still the same. The implementation does not really matter. If anything, it gives you job security.

EDIT2: if what the article claims does indeed happen, it's not clear whether you will have to change the algorithms. The idea of bringing memory and processor closer is not new at all. Your CPU has level-1 and level-2 caches. It's all transparent though. You can take advantage of them explicitly if you want to (with some SSE instructions), but it's not required. It could be done the same way.