Originally posted by jbirney
Thats the point I was trying to make. By sticking to one simple paths things are easier....meaning less transistors and easier optiomizations (remember jc said that scheduling for the NV30 was tricky). I am not saying either is better. Just saying that wiht every engineering choice is a trade offs. they both have "good" and "bad" parts...
Yes, yes, I figured that out.
But what I meant to say is that the calculations are always FP32. There were rumors that calculations were actually done in FP16 when you asked for FP16 and that this was two times faster, and that registers were just a small part of the problem: it's actually exactly the opposite.
FP Calculations are always done at the same speed, and the only difference is with register usage...