What is the process called that uses FP...

[JWMiddleton]

Hi TeAm, I have a question that I'm sure this group can answer...

What is the process called that uses FP engines of video cards to speed up calculations? I remember this being talked about a few years ago where a Radeon card (1900xt maybe) was being used to greatly speed up some distributed computing projects. Did this ever work out and did other number crunching application ever take advantage of it?

Thanks, John

 

[JWMiddleton]

Looks like I finally found it. Here is an article: http://www.pcper.com/article.php?aid=745

 

[Ken g6]

Hey, I remember you, but I guess you haven't been around here recently. Milkyway certainly does. Other projects, including the CUDA and OpenCL clients I wrote for PrimeGrid, use integers, but not floating point. They definitely speed up calculations!

 

[JWMiddleton]

Hi Ken, Thanks for the reply. Since you have written code for both CUDA and OpenCL, then you should know a great deal about the subject. Do they use FP or Integer math? I could have easily gotten it wrong.

 

[Ken g6]

Either or both. It's true that GPUs are generally designed for single-precision FP, but 24-bit integer math is easy to derive from that, and 32-bit integers a little harder but still doable. I'm doing 64-bit integer math with 24/32-bit instructions. Some cards support double-precision FP too, but usually much slower than the single-precision version.