Intel mulled buying Nvidia, ATI

FOR THE THIRD part of this exclusive interview with Intel senior vice president Pat Gelsinger, we talked about GPUs, how Intel considered buying ATI and nVidia, the OLPC laptop and the Tablet PC. But I started this latest leg of the epic journey asking about the appeal, or otherwise, of dedicated processors for specific tasks, such as Azul Systems’ monstrously multicore systems for accelerating Java…

“Accelerators are very interesting although there’s nothing new in this,” Gelsinger says. “We had media coprocessors, the math coprocessor with Weitek… [Firms like Azul] have fairly well defined market niches. The only [dedicated processor] that has really had sustainability is the GPU. We’ve had integrated GPUs for some time. Now, [we are moving] to bring the memory controller and GPUs on board [the CPU].”

Gelsinger flatly denies that Intel would like to see the death of the dedicated GPU and replace it with all graphics processing handled on the CPU, but he does have a problem with some of the wisdom being spouted about the role of GPUs and rival CPU architectures.

“ATI, AMD and nVidia have said, ‘We have a lot of floating point right inside our GPUs, why don’t you take advantage of that?’ That’s a different discussion and it’s a major software port. Why did Cell fail or the Emotion Engine? Because the programming model is awful. Graphics is such a driver of performance that we don’t see discrete GPUs going away, but the majority of the market is now integrated. Not everyone needs global illumination. The majority will be satisfied [with on-CPU graphics] but there’ll always be gamers and people with a visualisation focus.”

DAAMIT (a nickname much admired by Gelsinger, by the way) was created because AMD worked out that it needed to buy ATI for its graphics computational know-how. Did Intel look at doing the same? The answer is, ulp, yes – even if Gelsinger says this was only part of a standard strategic evaluation process on the road that turned out to be heading to Larrabee.

“Pre the ATI acquisition we looked at it pretty closely and said, ‘Should we just go buy ATI or nVidia? One issue was that we didn’t know if we could because, if number one buys number two or three, what happens regulatory-wise? Also, we weren’t clear we needed to. The key transition [we’re going through now] is in the graphics programming model. The issue [GPU makers] have is making the pipelines more programmable, and we have the most programmable model on the planet - IA.”

OK, two more quick ones: is Nicholas Negroponte’s brainchild, the OLPC laptop, a threat to Intel because it is innovative and suggests the possibility of cheaper commercial products?

“The OLPC work is innovative and incomplete at the same time,” Gelsinger cavils. “They say, ‘We only support open-source software’. Well, the majority of Classmate PCs [Intel’s nearest equivalent to the OLPC kit] we’ve sold have been Windows, not Linux. Engineering platforms takes time. Mesh networking, reflective displays… it’s oriented to a certain class of business model. At the highest level, ‘Go, Nick!’, but exactly how useful will it be? That’s hard to say. There’s also the Eee PC from Asustek. It’s a cheap PC, it’s cute.”

What about the Tablet PC – is Gelsinger seeing signs of a change in customer feelings? Negative.

“Not really. When it came out I was sceptical over the slates. I thought the usage model would be a mixed-mode device. There are limitations of battery life which will help the category [but] the OS and application integration are still pretty clumsy. Even if you have the perfect device I don’t think it will be [hugely popular]. The UMPC is the one [category] I really think will take off.”

Tomorrow: we’ve saved the best for last as Pat Gelsinger gets his teeth into the big questions. Tune in to find out how Intel got fat and lost ground to AMD, whether the Wintel axis is intact, what our man’s personal and professiona l ambitions are, and the day he kicked Mike Magee. µ

See Also
Pat Gelsinger says chips have to keep getting faster Part 1
Pat Gelsinger on changing computing models Part 2