For my current involvement of taking the Udacity Self-Driving Car Engineer Nanodegree, I'm currently using a 750 Ti SC card - which was what I had in my current

For my current involvement of taking the Udacity Self-Driving Car Engineer Nanodegree, I'm currently using a 750 Ti SC card - which was what I had in my current workstation at home. I've found that it performs adequately (so far) for what I am doing. I'm running Ubuntu 14.04 LTS, and I installed CUDA and CUDANN (and other junk) to run Python/TensorFlow with GPU support. That said, I'm currently working on the behavioral cloning project (teach a virtual car to drive using what we've learned so far), and given the amount of data we'll gather (10s of thousands of frames we have to process) to train our models - I tend to wonder if I will need to upgrade. Fortunately, that won't be too difficult for me - I have both a 960 and 970 waiting in the wings; they were cards originally destined for a mini-itx gaming machine build that I've never gotten around to (hmm - maybe I should just build it for my next workstation instead). Also - 960s and 970s are dropping in price like a rock, since the hardcore gamers are moving to the 10xx line; while they may not have as many cores or as much speed, they are still very capable cards for ML purposes. I've also been considering getting a multi-pcie slot server motherboard (maybe with dual xeons or such - these mobo combos are actually pretty cheap on ebay), and dropping in one of those cards, then later adding on more GPU cards - strictly for an ML box...