[rescue] emulation fun

[nate at portents.com]

> (OK. A GPU is even more specialized then a vector CPU like the Cray 1
> and a GPU does only single precision. But still...)

Perhaps you're thinking back to the ATI RV670 (aka Radeon 3870) when it
introduced FP32 filtering support to simulate FP64, at half of the
performance in theory compared to single precision operations.  However
NVIDIA introduced native FP64 units on it's GT200 series (the GTX280 and
derivatives) and ATI also has native FP64 in it's RV770 (4800 series).

(Aside: there's been a bit of obfuscation on ATI's part regarding
marketing their GPUs, saying the 4850 and 4870 have 800 SPs, when really
they have 160 SPs made up of one FP64 and four MADDs, which can't be used
independently, so things like Folding at Home for ATI GPUs only use those
FP64 units and therefore only really have 160 SPs their disposal not 800.)

Going forward, I expect with the release of OpenCL and support from NVIDIA
and ATI that people will really be able to tap into GPU power, and GPUs
are only going to become more and more generalized.  Intel looks to be
going more the Cell route with Larrabee.  MacOS X 10.6 will be very
interesting with it's focus on 64-bit performance, multithreading, and
OpenCL - I could see a sufficiently high-end Mac Pro giving a Cray a run
for it's money in a year or two.

- Nate