CPU Architecture is Dead

about | archive

[ 2005-June-07 15:23 ]

I find CPU architecture to be a deeply fascinating topic. I love thinking about the performance/cost/power trade-offs that are inherent to CPU design. However, I believe that CPU architecture is a dead art because soon, all CPUs will be x86, which limits the possibilities for CPU design enormously. Apple's incredible switch to x86 is simply the most recent, most notable sign of this trend. It would seem that with its ancient instruction set, other chips like the PowerPC or the Itanium would have a substantial performance edge. However, that does not appear to be the case. The reason seems to be that economy of scale is more important today than superior CPU design.

Today, the x86 scales both up and down. AMD's Opteron and Intel's Xeon high-end processors dominate the Top 500 Supercomputers List from approximately the 40th position down. Intel used to say that the future of this high-end market would be with their Itanium processor. However, the Itanium has been a multi-billion dollar disaster. Processors like Centaur/VIA's C series' and AMD's Geode are starting to inch their way into the embedded market, as seen by the popularity of embedded PCs made by companies like Soekris Engineering and PC Engines. Transmeta was founded on the idea that superior CPU architecture could beat x86, but in order to make people even consider their chips, they had to build a complex x86 compatibility layer. Their risky strategy failed to pay off: Intel's mobile chips obtained significantly better performance using a more conventional design.

I am exaggerating the situation. CPU architecture will continue to be important, however the instruction set will dominantly be x86. Other instruction sets will only be found in parts of the embedded market where power and cost are critical, and compatability is a non-issue. Today, with Apple migrating to this old, familiar ISA, every major computer vendor supports x86. It is hard to compete with that kind of volume.