Created: 02 Feb 2005
Two recent articles on future processor architectures have fallen into my inbox in the past couple of weeks.
Herb Sutter has written an interesting piece for Dr. Dobbs Journal entitled The Free Lunch is Over. He discusses the plateau in processor speeds that we have currently attained and how future increases in raw processing power will come from hyperthreading and multicore processors, which require software to be written with concurrent processing in mind. And concurrency is hard.
IBM have their own idea of which architectures will dominate in the future. Here is a description of the cell architecture, using information gleaned from IBM patent applications. I cannot vouch for its accuracy; for one thing it predicts that Cell will kill x86, We’ll see. There’s a significant amount of inertia in the computing world and x86 has beaten off many challengers.
It seems that for once, software will be forced to use a processing model closely aligned with the hardware. The flipside of that is that hardware designers will need to pay attention to software engineers, or we’ll end up with a bunch of extra hardware features that no software uses. Like many of the instructions that Intel have added since the 486.
I’d like to see more recognition of the fact that applications are often bound by I/O performance. Even RAM is too slow for modern processors. We’d see some great performance benefits from doing away with mechanical storage. Surely this is one area where we can expect cheap gains in performance from the hardware.