Prime95 software for PS3/GPU?

Looking at the success folding@home has with PS3/GPU support, this would be nice for Prime95, too. This was already mentioned in some threads, however, the argument against it is "only 32-bit floating point supported", and the FFT that is used needs 64-bit floating point numbers. But there are 2 ways to work around this, or am I wrong?

1. Use integer-FFT instead of floating point FFT.

Quote from the "The math" side ( http://www.mersenne.org/math.htm ):

Although GIMPS uses a floating point FFT for reasons specific to the Intel Pentium architecture, Peter Montgomery showed that an all-integer weighted transform can also be used.

Does a "normal" version of Prime95 with integer-FFT exist already?

2. "Emulate" 64-bit floating point support.

64-bit integers can be emulated using 32-bit integers (using some shift operations), the same is possible for 32-bit floating point numbers (there are also some papers around how to do this at least for NVidia GPUs). However, this would definitely decrease the speed. Anyway, if the theoretical 100-500 GFlop/s would decrease to 10-50 GFlop/s, it would still be a lot faster than the CPU alone.

Greetings,
schnaader

[jasonp]

Originally Posted by schnaader

1. Use integer-FFT instead of floating point FFT.

Does a "normal" version of Prime95 with integer-FFT exist already?

2. "Emulate" 64-bit floating point support.

Regarding number 1, nothing much has changed since this thread. If your CPU has a pipelined 64-bit multiplier then you have a shot at an all-integer transform that's fast enough. What's the widest multiply you can do on a Cell engine? If it's like the altivec PowerPC architecture, you can do four 16-bit multiplies at a time. Plus Montgomery's algorithm requires access to carry bits in order to be efficient.

Regarding point 2, emulating 53-bit floating point with 24-bit floating point will still limit you to an 8-bit exponent, and this may not have enough range to handle an FFT with a million elements. Plus of course it's 10x slower. FFTW in single precision on the Cell, with seven engines, manages 20 GFlops for moderate size transforms. I doubt you can achieve that rate when the floating point primitives are 10x slower.

Edit: the biggest hurdle is convincing George Woltman (or whoever is interested in this) that there will be enough Playstations sold to prime number enthusiasts to justify a port to a completely different architecture. I sincerely doubt the number of consoles is going to rival the number of PCs anytime soon.

jasonp