I originally came across this thread, which made me believe that the included OptiX denoiser model was only tied to a driver version.
However, in a few recent experiments I ran, the denoiser produces slightly different results on different GPU models which have the same driver version. I’m fairly confident this was not the case on an older driver version (we recently upgraded from v510.xx to v515.65).
If we wanted to produce exactly the same outputs on two different machines, which parts of both systems/environments would need to match exactly?
Different GPU architectures make use of different kernels. Means even with the same training data, the calculations could be slightly different, similar to fused multiply add and separate multiply and add instructions.
Everything must match to get exactly the same results. That should have always been the case across different GPU architectures.
Thank you, that makes sense!
We weren’t 100% sure on which machine we had generated our “reference” de-noised outputs anymore, hence the “fairly” in my original message.