Ask HN: How to estimate compute requirements for Machine Learning algorithms?

Has anyone developed formulas or rules of thumb for estimating CPU, memory and time requirements for running various Machine Learning algorithms, with respect to training and/or prediction–classification? I realize that implementations may vary and that a variety of parameters can be tuned, depending on the algorithm, thus affecting the requirements, but I'm interested in ballpark estimates.