Are Programmers Up to the Task?
- 2
- Add a Comment
A not so unexpected announcement from Intel comes today, telling us what we already knew – the speed race, in GHz, is over. The chips of today are close to their limits already – there will be no amazing 6GHz chips just over the horizon. Oh sure, there are those apocryphal tales of old style Pentium 4s hitting 5 GHz under nitrogen cooling. True? Probably. Possible for you and me? Not likely. Unless some gains are made in room temperature superconductivity, we will see all new speed increases from parallelism.
Massive parallelism. Or so states Intel. Wait a minute – isn’t this what nVidia’s CUDA initiative is all about? Yes, but that’s another story that Intel wants you to forget.
What Intel wants the public to know right now is that for this to work, all your favorite applications are going to have to be extensively rewritten. Alright, you say. Well, the problem here is cost. It’s going to cost a lot, because the way that programmers think is going to have to change.
The writing will be changed. The verification will also have to change. It will all begin with the compilers used, and those will have to be optimized for the number of processors expected – if the compiler is designed for 12 cores – 12 threads of execution, and someone wants to use it on a 16 core system, those extra cores go wasted.
Beyond that, there are many jobs that don’t respond to massively parallel computing – 1 thread must complete before the next, and no amount of branching and prediction will help.
As astounding as the resources of Intel are, nVidia has this working now, at least to a much higher degree than Intel. The only other possible competitor at present would be the owners of the Cell processor.
With nothing in the Intel road map according to its tick-tock time frames, it will be a horse race to see if it is nVidia or Intel, or possibly IBM, that gets the massively multicore processor to the average desktop first.
-
Technorati Tags: multicore processor technology - Intel - nVidia - IBM - Cell - CUDA - threads - pipelining - branching - prediction
2 Comments
George
July 2nd, 2008
at 4:20pm
I’ve also heard that intel was experimenting with light as a viable source for processors? needless to say, it would be very expensive, but also much faster. I’ve also heard that they might be experimenting with moving actual atoms themselves. does anyone know if this is true?
the oracle
July 2nd, 2008
at 9:33pm
George, what you are talking about is farther away than what I speak of. Also, the same parallelism will apply, as there will be a limit to every serial process in terms of speed.
The major thing is that the way in which the problems are solved will require a much different process.
Thanks for the comment.