Intel's chief technology officer Justin Rattner isn't buying AMD's claims that the chip titan swiped key technologies from AMD and applied them to its own forthcoming processor designs. In fact, he says, many of the changes the company has in store for new chip architectures like Nehalem and Larrabee have been in the works for years.
During Intel Developer Forum last month, Wired News sat down with Rattner to discuss why "enthusiasts" are so important to chipmakers, what he thinks about tri-core processors, and why sticking to older architectures helped Intel get a leg up on its rival for close to a year.
Wired News: Paul Otellini's IDF keynote theme focused on moving from the extreme into the mainstream -- aiming products at the high end and letting those features trickle down to the average consumer. Are you still using this MO in the multicore era, given the fact that quad-core still accounts for 2 percent of the market after close to a year?
Justin Rattner: I think it's pretty traditional in the computer industry that you tend to innovate at the high end, where developers and users are more willing to deal with the unknown. I spent a decade in high-performance computing, and the users of those machines wanted to be at the bleeding edge.
That was just the name of the game. We're more than willing to perform heroic acts of programming to eke out a little bit of performance. That's nothing new. You can go back into the history of computers and they always seem to start doing something incredibly demanding for a very narrow audience and then broaden out over time.
WN: I bring it up in the context of the generally accepted fact that there's a wide gap now between the capabilities that multicore hardware offers and the ability of the software community to capitalize on that.
Rattner: Well, did you ever think how long it took Microsoft to deliver a 32-bit operating system? We had (a 32-bit processors) something like a decade before Microsoft figured out what to do with it. I think that's reality.
If you think you can wait around and the software developers will suddenly come to you and say, "Wow, I have applications that really need four cores, could you just build one?" and you'd say "Oh, OK, we'll get right on it." That's never worked. We've always had to put these architectural features in first, whether it be the number of cores or virtualization or a lot of security things. Eventually programmers discover these things and realize that they are actually pretty useful and they start building to the architecture.
But I think hardware always has to lead and, yeah, it's a big gamble. You build a $4 billion fab and you hope sometime soon somebody figures out how to use this thing.
WN: AMD recently announced it will be releasing a tri-core processor. What are your reactions to that? The theory is they're releasing it at least in part because of the small market for quad-core chips and that it's going to take some time for people to want to move up.
Rattner: I wouldn't make that much of it. This is a yield-improvement technique, plain and simple. IBM and Sony with their Cell processor -- they have eight (processors) on that that chip, and they said, well seven is the actually the number and one is a spare, or one is dead. I'm never quite sure whether there's a dead one or not. It's just like memory chips today … there are thousands of spare memory bits that are there (to ensure sufficient) yield.
So, yes, (AMD has) a four-core product. I'm sure when they looked at their yield losses, they said, "Wow, we can offer a three-core version of this if one of those cores are dead or slow or whatever it turns out to be."
In terms of software, there's software for one core and then there's software for multiple cores. It's not like, oh, we have a three-core problem but not a four-core problem.
WN: According to AMD, this is a specific product for a specific time -- for this catch-up period before quad-core chips take off.
Rattner: Either way, this is strictly transient because when we start thinking about tens of cores, then we actually start thinking about how we won't build 32, we'll build 36 and we'll have those spare cores so when cores fail we fail down to the spec, right?
When cores are relatively inexpensive, they represent a few percent of the total area. Yeah, throw in a couple extras and you'll increase your yields. You'll guarantee the customer a minimum performance level and they don't have to worry about their customers coming back and saying: "Gee, this seems to be slowing down. What's going on?"
WN: What about architecture, specifically native quad-core? AMD is claiming that new features like QuickPath are analogous to their own Direct Connect architecture and HyperTransport. They're casting your move as validation -- that they were right all along.
Rattner: IBM, in the heyday of the PowerPC, made this big to-do about using copper as the metal that runs between the transistors and hooks them up. Our view was when we need copper, we'll have copper. We didn't feel we needed copper. And in fact we used this analogy of traffic lights. Copper allowed the cars to go faster between the lights. But the lights, which represented the transistors, weren't changing any faster.
And of course the proof of the pudding was that IBM, despite copper, never built a PowerPC that was even half the frequency of the leading Intel processor. So gee, if copper has this magic property of making faster processors, how come your processors are less than half as fast as Intel's processors?
Now, we did get to the point where the transistors got fast enough where we said copper makes sense and we introduced copper. We actually had the copper processors kind of sitting on the shelf just waiting for the point when transistors caught up.
We said the same thing a year and a half ago when there was a lot of competitive noise about AMD's integrated-memory controller. We said look, the design of these things involves thousands of tradeoffs. We felt we could achieve the requisite performance levels without an integrated memory controller, which adds cost to the processor. We said when we get to the point when we feel (an) integrated-memory controller is required to achieve the required performance levels, we will integrate the memory controller.
The same thing goes for the point-to-point links between the processors. In fact, I like to point out that Intel was able to do a quad core -- putting two dual-core dies in a single package -- because it had a front-side bus architecture. We could just tie the front-side buses together internal to the package and drop quad core and do a dual-core socket. No change.
So we were able to get quad-core to market a year ahead of AMD. It was not an option for AMD. And I believe if it had been an option, they would have done it. But it's not an option in their link-based architecture -- in that HyperTransport architecture. So they were basically not players in quad-core for almost a year.
You can talk about architectural elegance, but in fact we've shipped millions of qua- core chips. That small percentage of the market represents millions of processors. As a result, we've got eager users and enthusiastic users and we've got motivated software developers.
What's the old adage? We will sell no wine before its time? Well, we will sell no processor before its time. We think the Nehalem generation just called for those features and they're in there.
