Intel has formed a new design and development group for 'SoC enablement'. That's System on Chip – in other words, making processor chips that include all of the other electronics to do a particular task. So far, SoC has been mostly used in appliances, electronic products such as set-top boxes or routers.
It's a huge area, but one that's an uncomfortable match to Intel's preference for making extremely powerful processors that can do anything. As soon as you start to include more specific functions on a chip, you're making it more attractive for a particular market at the expense of all the others; you're also limiting the support it's going to get from development tools and the community.
Indeed, Intel's previous efforts for SoC have not gone that well – it's had some successes, but rather more failures. The 80186, for example, combined the 8086 core with some improvements and a lot of the support circuitry that normally surrounded the processor in a PC – but at the expense of compatibility. PC makers who used it, suffered, although it's still around as an embedded part. More recently, the Timna chip combined a Rambus memory controller and video controller with a Celeron core: as Rambus prices remained stratospheric and customers remained brutishly uninterested, it got canned at the very last minute. (The Israeli design chip behind Timna went on to make Banias, the heart of Centrino and the guiding light for nearly everything Intel's done well since, so it's true what they say about some mistakes being instructive. Even very expensive ones.)
Yet here we are, nearly a decade later, and Intel's having another go. There are obvious commercial reasons – AMD has been successful with its more SoC-ish approach – as well as practical ones to do with having to tightly couple high performance components with the CPU in ways that are increasingly difficult to do off-chip. But the real driver, I suspect, is the changing face of processor design as a whole.
In 2008, we're going to see Larrabee, the first of Intel's many-core chips, creep into the light. As we know from our experiences with double- and quad-core chips, there's a severe law of diminishing returns in just ladling more cores into the same package. Some things get a lot faster: most things don't. And the more cores you have, fewer situations fall into the former slot while more end up in the latter. That sounds a lot like the same problem you get making an SoC chip general purpose – and since the only way forward is in increasing core count, it sounds a lot like Intel's going to become an SoC company whether it likes it or not.
There is little future in sticking to one generic many-core design. Larrabee isn't just a chip, it's "...a graphical computing engine which will be a massively parallel, high-power, high-performance product line moving into the most information intensive segments of the market place.", according to Intel CFO Stacey Smith, while at the other end of the scale Silverthorne is a very low power x86 core that's going into SoC chips like Moorestown – designed to hit extremely cheap laptops and embedded devices alike. Larrabee will also have to include lots of specialised circuitry in its various incarnations, and be tuned for each – again, that's much more like ARM than Pentium. You're only going to get the competitive edge if you can tune your product for more – usually smaller – markets. Oh, and make the chips cheap enough to matter.
The big question is; can Intel make the transformation? Its entire chip production system – one of the wonders of the modern industrial age -- is based around doing a few things on a huge scale very well indeed. "Copy Exactly" is the magic phrase. But if it is to start making many, many more variants of complex systems built around an increasing number of building blocks – each complex enough in their own right – how does it build in that flexibility while keeping the economies of scale that drive its business model? Silverthorne is tiny – you can fit 2500 on a single wafer, as opposed to 400-500 Core 2s – but what if you want to make fifty variants of a Silverthorne-based SoC on that wafer, let alone decide what they should be, test them, package them and sell them?
These thoughts must run through the entire company – marketing, finance, research, development, design, production, sales, software developer support – and in every case, the changes required will be profound. Other companies are far further down the line; excluded by Intel from the PC processor market, they've had to adopt much more flexible design and production ideas and have a lot more experience - and intellectual property – in this area than Intel. (Expect some acquisitions. And you wondered why AMD bought ATI?)
Lots of questions. Intel's long-term success will depend on keeping its disciplined company philosophy in place while applying it in new and increasingly radical ways.