… make Word, Excel and PowerPoint, just as you know them today, 100 times faster, would it be significant for you? The answer is probably not. If I told you that Word became able to do perfect dictation, would that be interesting? It might be. And so there are some qualitative changes even in traditional apps that I think can be brought about by this computational capability, but only if we are able to write the code in a way that really exploits it.
One of the things you talked about is the fact the supercomputing industry has been tackling this for a long time. Microsoft recently scaled Windows to handle some of those [cluster computing] tasks. Is one of the reasons Microsoft got into that market the fact that you saw the whole computing industry is going to look more like cluster computing does today?
In part it was. I was one of the big proponents of us moving into the high-performance computing space. My heritage is in that space, too. And I do see some incredibly strong parallels between what the supercomputing industry moved to in its last 10 to 15 years of evolution, and what I see happening in miniature at the desktop now.
I think there are some lessons that we can learn from our high-performance computing efforts that probably will help us, as every desktop has that level of capability or that architecture at least.
One of the biggest changes is that if you're going to scale the core count very high, the memory system architecture has to change; the traditional bus-oriented memory systems of personal computers won't scale to happily feed the hunger of all of these many processor cores, and so that brings with it another challenge or set of challenges.
When you talk about a very large number of cores, do you have a sense of a few years from now what types of systems you'll be looking at?
Well, if you see the progression now, in the last couple of years we went from one core to two cores, and we've got four cores coming today. I have no reason to believe that the core count won't sort of stay on a trajectory that matches the transistor doubling.
The limiter then becomes "What is the architecture?" If they're completely disjoined, every one is unique and they have their own pins to talk, then you could say, well, you could just keep going. But, in fact, you run out of other physical things before you're probably ultimately going to run out of cores.
Vista Upgrade Blog
Grappling with the OS
How is the switch to Vista affecting your workplace? Take a look at our new group blog and share your pain and praise.
You have to figure out how you talk to them. That means that the input and output on pins physically will be different. How you get power into the chip and how to cool it continues to be a challenge. But I think certainly over the next decade I have no reason to believe you wouldn't see core counts that were north of 50 in these large chips.
One of the things that you've talked about is that all these cores give you the luxury to have the PC speculatively do some task that you might or might not do.
That's right, power to burn. One that I think is the tip of the iceberg, this thing we did in Vista with SuperFetch to accelerate loading of the next app you might run, required us to develop a model of your behaviour and the time of day and what you've been doing and how many resources are there. But by speculating on what you're most likely to do next, we're using basically idle capacity of the machine to make it appear to be yet more responsive on some of the long latency tasks. I think that that kind of thing can be expanded quite dramatically.
I mean, if you think how computer plays chess, right, I mean, what it does is in any given amount of time it runs through all the permutations and combinations of the moves that can be made from a current position, weighting them using…
