GPU requirement for convolution neural network.

Matlab example (https://www.mathworks.com/help/nnet/convolutional-neural-networks.html) suggests that "using a GPU requires a CUDA®-enabled NVIDIA® GPU with compute capability 3.0 or higher" is recommended for CNN. Wiki says that 3.0 series is kepler architecture, CUDA 6.0 is Pascal architecture. Also price of their models varies wildly (from 500$ to 6000$). I wanted to know that is there optimum GPU series that I should buy? Can anybody share some first hand experience.

Answer 1

[Joss Knight]

Current best value for deep learning, image processing and computer vision applications is the GeForce GTX 1080i or the Titan X (Pascal). For scientific computing you'll want a Tesla card with the Tesla K40 being good value as a desktop card and the P100 being the best server card currently available.

The distinction is about whether you are primarily using single precision or double precision data. The GeForce cards are good value but absolutely terrible at double precision computation, so rule them out if you're doing system modelling, finance, or something else that needs double precision maths.