I went to Mike Muller’s keynote at ARM’s techcon3. He started with an interesting retrospective on ARM. They have shipped 15B units (4B in 2008 alone). They have 20+ processor cores, 600+ licensees. In the next 3 or 4 years they will ship another 15B units. It’s not far off to say that “almost all” microprocessors are ARMs (by unit count).
Smartphones are one big driver now. Basic and enhanced phones have peaked and volumes are actually declining, but smartphone is growing fast. For example between 2007 and 2008 mobile phone browsing increased by 30%. It is clear that smartphones are going to become the dominant way of accessing the internet (they already are in Asia).
If you look out the next 10 years then in one aspect things look rosy. In 2014 area will have increased by 4X from today, by 2020 increased by 16X. Clock frequency is still going up, although not dramatically: up 1.6X by 2014 and 2.4X by 2020. Volume production of 3D chips (with TSV, through-silicon vias) will come along in this period leading to “motherboard in a package”.
But unfortunately power is the fly in the ointment. Power for given functionality at a given clock speed will decrease to 0.6X by 2014 and 0.3X by 2020. That’s nice but nowhere near enough. Run the numbers: 16 times the area at 2.3 times the clock rate and 0.3X of the power means that only 10% of the chip can be used if the power budget remains fixed (which largely it does due to heat and battery life). So complex power architectures, power down blocks, adaptive clock frequency and so on are going to be essential going forward. Design is getting more and more complex.
The titanic battle is, as I’ve said before, between ARM and Intel’s Atom. Intel were cheeky enough to bring a huge truck and park it outside the ARM conference, which Mike had managed to get a photograph of into his keynote only an hour or so later. Atom has two advantages over ARM: it has binary compatibility with Windows and the PC, and Intel’s manufacturing is second to nobody. My belief is that as more stuff moves online and more content is accessed through smartphones rather than PCs, then binary compatibility will be less and less relevant. Further, Intel won’t be manufacturing most Atom-based SoCs, so I’m not sure either advantage is strong enough.
Mike finished with a Lucite block from a microprocessor forum ten or 15 years ago. It contained all the new processors being announced at that forum (back then, that was where everyone announced new processors). There were 11 different architectures. Of the architectures in the Lucite just four remain today: ARM, Intel x86, MIPS and SPARC. Mike is pessimistic that MIPS and SPARC will survive long-term (and I agree) leaving just two, Intel and, of course, ARM.