Specialized chips

[jasong]

What do you guys think of the idea of computer manufacturers making specialized chips for specific things, like number-crunching? I've asked in other forums and people keep saying it'll cost millions of dollars. I'm not saying this'll make a difference, but I'm not talking about severe changes, I'm talking about simple changes. For instance, smaller or larger cache sizes, making chips that are specifically made for number-crunching without necessarily optimizing, etc.. I realize these chips would be more expensive, but would they be more expensive for the specific task they're made for?

Comments, please.

[wirthi]

Why would you want to buy such a processor.

I mean, I don't mind my (necessary!) computers run 24/7 instead of 12/5, the ew extra bucks for electricity don't really count.

But I wouldn't buy a computer that was specifically made to crunch a singe Distributed Project (or a type of projects) just to cruch faster if the CPU had no other use.

I mean, that's sensless, this distributed thinggy is not about producing new computers to run the projects on but to use the otherwise wasted power of computers that have other uses besides distributed computing.

Just my 2 cents.

(Well, if you get AMD, Intel or someone else to produce a tiny minicomputer that is cheap AND is fast on distributed projects I could change my mind but I guess that's quite an expensive project ...)

[Scoofy12]

well, for starters, a microprocessor IS a device for number crunching.
so what are we talking about here? a specialized CPU is what it sounds like when you talk about extra cache, etc
so what would you optimize for? even among your "number-crunching" DC projects there is huge variety of workloads. DF, for example, spends most of its processor time chasing pointers, which is more like office applications or everyday user programs than floating-point intensive scientific calculations that you are probably thinking of when you say "number crunching." (so for these, you probably won't get any better than P4s/Athlons). if you mean adding more floatin point hardware perhaps, some more FP-intensive applications are quite vectorizable (as evidenced by good performance on mac/altivec), while others are not.
so different improvements would probabl only benefit a very narrow range of
applications, in a sector which would already have very low volumes, saleswise
for example... guess what? there already exists an extremely fast FP processor, which also has pretty poor integer/general purpose performance.... it's Inte's Itanium. actually there are other such chips as well, but youve probably heard of this one. it's really expensive. why? mostly because of low sales volumes. the more-specialized hardware is even more expensive. you're talking about something even lower-volume than that. and for a variety of reasons, the fewer you sell, the more expensive per-unit the cost is.

and of course, there is no "just specifically made for number crunching without optimizing"... say for example you want to add more floating point hardware. in order to translate this to faster execution, you need to be able to issue more instructions at once into your functional units. this means looking farther ahead for instruction-level parallelism, which basically involves increasing the complexity of the entire processor pipeline, both conceptually, and in transistors, which means major design expense. you can't just plop more functional units in there, you have to keep them fed.

as for cache sizes, they are already increasing cache sizes as much as they can, and despite the fact that bigger caches are slower, it's still better that they be big.