nAo: TMTA used binary translation to provide x86 compatibility. Their first chip never made it out the door and Crusoe was kind of crap. Their 3rd chip looked better, but by then the company was a goner.
If they had ever gotten a lot of traction, they would have gotten sued by Intel...but they didn't.
To me it's a technology marvel, just like the first DRAM chip. Not very practical, but always amazing to look at. I actually bought a small notebook (very similar size to current netbooks, by the way) with Crusoe CPU. The general performance is unfortunately quite lacking, about the same speed of a half clocked Celeron (i.e. a 600MHz Crusoe is roughly the same as a 300MHz Celeron).
However, its compatibility is really amazing. I used that notebook as a normal one and I never encounter any incompatibility. It just behaves like a normal x86 processor. It's hard to believe that all x86 programs are run through emulation at a pretty reasonable performance. It's even more amazing after the underlying architecture was revealed to be a very simple VLIW instruction set.
It's binary translation, not pure emulation. Big difference. Infrequently encountered code was emulated, commonly executed code was BT'd.
Of course, from the outcome we now know that a pure software solution is not the way to go. Many of Transmeta's power saving tricks are not related to their "code morphing" approach, and can be used on normal hardware based x86 CPU as well. Actually many of these tricks are used in current mobile processors and resulted in some patent lawsuits between Transmeta and Intel, which Transmeta earned a settlement with Intel.
TMTA never filed any of the patents that they nailed Intel on. They bought several patents from a 3rd party, which Intel was found to have violated. In fact most of TMTA's own patents proved to be worth not much.
Binary translation is very interesting, but one of the big problems was that you could either translate the x86-->VLIW or you could execute your translation, not both. Whoops.
Also, the code bloat was awful for windows. They really needed much much much bigger caches...1MB probably wasn't enough (and that was their L2, L1I was smaller).
DBT has it's place and is a great technology. Intel is basically doing minimal DBT in hardware with their stack engine and the macro-op fusion. DBT can be a great way to support some legacy crap (e.g. we don't need x87 hardware any more).
However, TMTA was founded on the idea of building a faster, simpler x86 using VLIW...based on a bunch of simulations done on SPARC. That plan never materialized and became 'lower power x86' - not a bad business plan, but much more difficult.
Intel proved that it's a hell of a lot easier to take a P3, reduce frequency and voltage and use your best bins...rather than designing a brand new processor. And when they did design a brand new processor for mobile, that was the end of TMTA. TMTA made a bunch of quite fatal mistakes along the way...but if they had a gravestone it would say 'thanks for the Pentium M'.
David
