Axiom Home Page Forums General Discussion The Water Cooler Bad math for video

When I was doing vector programming on the Cell processor I noticed that IBM was making a big deal about the Cell being able to perform double precision math while all GPUs at the time were limited to single.

But there was a 4x speed hit for using doubles over single precision floats. So the extra precision was there for people who needed it, but since I was processing raw data from camera sensors, that at most had 24-bits of precision (but the noise floor really only yielded 18-bits of accuracy), the single precision was enough.

Though the point is I could get the extra precision if I needed it. Chips that trade speed for accuracy will be extremely single purpose. There's no way to turn off the design flaw when you don't mind trading in the speed, like there is with traditional design.