The paper presents a broad strategy for dealing with nonlinear dependency. It claims to be up to 80x quicker in a calculation, and, more significantly, it exhibits linear dependence on seismic size.
The authors provide a comparison of CPU-based analytical and GPU-based neural network attributes calculation. Even though it's not a fair game, but it leaves thoughts to consider.
The publication: João Paulo Navarro et al.  Real-time seismic attributes computation with conditional GANs
Links are in the comments.
The link for the paper
The link for the LinkedIn post