Geforce 8800

[ATH]

I know the question about Prime95 on GPU's have been asked many times before. But the new DX10 cards Geforce 8800GFX and 8800GFS have a new feature called CUDA (Compute Unified Device Architecture):

http://www.tomshardware.com/2006/11/...00/page11.html

[cheesehead]

Originally Posted by ATH

I know the question about Prime95 on GPU's have been asked many times before.

... and if you want to see any different answer than you've seen before, then the question you need to ask is, "Does this GPU offer what the others didn't -- namely, double-precision (64-bit) floating-point arithmetic?"

For L-L purposes, single-precision (32-bit) floating-point arithmetic doesn't provide enough bits to be worth the effort of implementing the irrational discrete weighted Fourier transform (IDWT) algorithm that provides the fast multiplications in Prime95. (IOW, George could do it, but all he'd get is a slow GPU implementation. Cranking up his old P90 would be cheaper because it'd use less electricity than the new GPU for about the same speed.)

Looking through the linked article, then also through "What Direct3D 10 is All About" at http://www.tomshardware.com/2006/11/..._is_all_about/, I eventually found, on page 4 of the latter, the statement "While everything inside Direct3D 10 hardware will be 32-bit (floating point or integer), there are several storage types. Below is a list of the data storage formats." That list shows widths up to only 32 bits.

So, alas, the answer (again, as with every other GPU I can recall, for the same reason) is that it will not be practical to port Prime95 to the new DX10 cards or to any future new graphics card unless it supports double-precision (64-bit) floating-point arithmetic.

(... or unless the GPU supports fast-enough integer arithmetic so that an integer-based transform method could rival the speed of the IDWT method. But I haven't seen that yet, either.)

- - - -

BTW, if you're saying to yourself, "but graphics cards don't need double-precision floating-point arithmetic for their intended purposes", then my response is ... "Aha!"

[RMAC9.5]

The Peakstream folks at http://www.peakstreaminc.com claim to be able to do 64 bit floating point math on ATI R580 graphics cards using C and/or C++. I know that George writes assembler code so this isn't useful to him but it might be useful to those folks who have created versions of Prime95 that use the C compiler for Linux. This web site also mentioned a no cost evaluation program for Linux workstation users.