I very much doubt that. This is why I dislike inflated claims by manufacturers so much - they are misleading, and often, when placed in context, inaccurate. Do you have a reference for this, or is it just a press release? No matter, I found it myself....

As I suspected, it is a distortion of the truth (a lie?) - apparently, what they mean by 128 bit is 4 independent 32 bit operations. Call me old-fashioned, but I expect the quoted data size to refer to the precision, not the portmanteau operand size.

And you wonder why I dislike marketing nonsense! I've yet to see a comparison of a Cell done by anyone without a vested interest in showing it off. In computing, it is very easy to optimise for a single processor architecture.

Don't believe everything you read, folks!

Thanks, Jord - I had never seen the Paranoia link before.

The GPGPU FAQ sums up a lot of the actual and potential problems with this approach.

Iterative approaches can be very, very sensitive to rounding errors. I don't know which or how many of the WCG projects rely on double precision - but being limited to single must be punitively restrictive. Floating point has enough pitfalls already, with numbers that can't be represented with any size mantissa, and unintuitive results to catch out the unwary.