Choose a graphic card to train SegNet for deep learning

I just started to learn semantic segmentation using SegNet. I'm in the stage to choose a graphic card between GTX 1060 6gb or GTX 1060 3gb or GTX 1050 Ti 4gb. I'm very budget tight so I won't consider anything beyond GTX 1060 6gb. I need suggestion whether GTX 1060 6gb is sufficient to train SegNet? My data base is small and it won't exceed 1gb of images. Or can I start from GTX 1050 Ti 4gb? I don't mind about slower training speed (I can leave my computer on overnight) but I'm a bit worried if 4gb is enough?

Let's say can I use GTX 1050 Ti 4GB for this Matlab tutorial

https://au.mathworks.com/help/vision/examples/semantic-segmentation-using-deep-learning.html

Thanks.

[Walter Roberson]

I am not familiar with SegNet; as I scan the description it looks to me as if it uses single precision, is that correct?

[Walter Roberson]

It looks to me as if the GTX 1060 is about 50% faster than the GTX 1050 Ti; https://www.videocardbenchmark.net/gpu.php?gpu=GeForce+GTX+1050

[Kyle]

I'm not really familiar with network precision. Yes, I understand 1060 is way faster than 1050 ti. But as a budget buyer, I'm wondering if 1050 TI 4GB can be used for segNet training as well? Because I cannot find hardware requirement for segNet.

[Joss Knight]

To answer the question on precision, all Deep Learning is single precision. This is also true of most image processing, although there are some notable exceptions (for instance imgaussfilt which by default computes in double precision for accuracy reasons).

Answer 1

[Joss Knight]

The GeForce 10 series will all work fine, they just have different capabilities and constraints. On a cheaper GPU you may need to keep the training MiniBatchSize down in order to lower the stress on memory and compute power. I've seen quite a lot of problems with kernel timeouts and other issues posted to Answers on the 960, 970, 1050 and 1060, but I do think most of the problems can be dealt with that way.