GPU utilization is high but memory usage is very low leading to subpar sampling performance

[sammosummo]

Would this model be a good candidate for GPU speed up:

Advice for speeding up large multivariate models

Hello all, I have the following model. Conceptually it is quite simple, but there’s a lot of data. Using the default options (i.e., ADVI followed by NUTS), it takes about an hour to run on my desktop Mac. I’d like to speed it up as much as possible. I wondered whether anyone has any suggestions for speeding it up, perhaps using some new features from 3.1, GPU integration, re-parametrizing, etc.? import numpy as np import pymc3 as pm from theano.tensor import dot with pm.Model(): # known…

In practice, these models will be ran on huge datasets.