Comment Articles

Are the old orders crumbling?

Tony Westbrook AnchorDesk

Published: 20 Jan 2000 14:17 GMT

The Crusoe is described as a processor whose instruction set is implemented entirely in software. Of course it is no such thing. Silicon is silicon is silicon. Software is software is software. You can't turn one into another any more that iron can be turned into gold.

No. What Transmeta has done, aided by its high profile alchemist employee Linus Torvalds, is create a new layer of abstraction between a very large instruction word (VLIW) processor and our applications. This layer is the part that has been executed in software rather than being burned onto silicon, and consequently it can be can be modified and developed to offer different personalities. Also, downloads can be provided online to fix bugs in the original's code rather as we are now used to doing with our PC's flash BIOS...

This extra layer of translation software is called code morphing software in Transmeta-speak. Another clever part of the implementation is that once an instruction has been 'morphed' it then resides in a cache and can be re-executed without troubling the VLIW processor at all. This is one of the techniques employed to reduce power consumption of the chip.

I think this soft approach is the right one to take in a world where we can all download bug fixes through a standard medium -- the internet. And we should not underestimate how complex a job it is to do either. Making the analysis about just how well its been realised will be the next job of many observers around the world. Interestingly, given the Torvalds involvement, but much more significantly the Paul Allen dollars, all the effort has been giving the first Crusoe chips an X86 architecture personality.

Of course, when you start doing emulations of foreign instruction sets, you must pay a performance penalty, and though the Transmeta CEO David Ditzel was slightly cagey about talking of specific performance, the impression given was that we were looking at a device that would offer marginally lower performance than current PIII chips. Once again labs like our own PC Magazine and IT Week facilities will be testing against these impressions as soon as they get hold of products -- within six months if Ditzel delivers on his promise.

Maybe the extra software layers were predictable enough given the history and personalities involved in the project. But there was something even more significant in the announcement...