Originally posted by jbirney
Actually from an engineering stand point is a great way to save on complexity. If you have one pipe thats FP24 all the time only, then its easier to manage that a FP16/32 or having seperate pipes. Its a comprimise and not a bad one at that.
Well, you know, I actually think the NV30 is really only FP32 - no FP16 calculations supported.
The only difference is temporary registers who can be either FP16 or FP32, and register usage is very important on the NV3x ( the least, the better )
So, there's really 4 pipelines who can do 1 FP32 op and 2 FX12 ops in parallel - but it can't do texturing at the same time as the FP32 op.