pymc3: parallel computing with njobs>1 vs. GPU

Question

I am trying to speed-up pymc3 sampling with parallelisation and I see only modest benefit.

I was able to decrease total running time from 25 minutes (njobs=1) to 13 minutes (njobs=6) on i7 MacBook Pro. Due to the fact that it takes about 4 minutes before pymc actually starts sampling, the increase is relatively small.

The question is - does anyone successfully using GPU with pymc3 and how much benefit can I get for models that take 6-8 minutes to sample? (My MacBook has nvidia GT 750M 2Gb)

Answer 1

I'm running Linux on an Intel i7-4930.

I ran a PyMC3 model that took 90 minutes on the CPU (utilizing all cores), but only took 18 minutes on my GeForce GTX 970.

So a speed-up of almost 5x.