The problem has been fixed now (2.78a) from the Blender side:
Well, I publish this here but I’m not sure if this should have been in Python Support.
I would like to discuss this because it is an annoying problem when it appears, because first, it is difficult to detect where the problem is caused if you do not know that addons can cause this kind of problems. And second, because if you have enabled many addons, then it is difficult to find the problematic addon.
So the problem is that if Motion Blur is enabled, these problematic add-ons make the memory gradually fill frame by frame without being cleared when rendering animation, until the memory is full and Out of Memory error message appear.
Right now I’m having the problem again:
CUDA error: Out of memory in cuLaunchKernel(cuPathTrace, xblocks , yblocks, 1, xthreads, ythreads, 1, 0, 0, args, 0)
but still I have not detected which is responsible addon this time (I have a lot enabled).
Here are some links where I found other people with the same problem:
So, why Blender allows an addon can cause this problem?. Is it a problem with Blender and python, or a mistake with bad code that add-on programmers are committing occasionally?. If this is a programmer mistake, Would there be any way to make Blender detect that mistake and avoid the problem?
I think I’ve found the culprit addon this time. It is “proxy_picker.py” part of Artell Auto-rig addon:
Auto-rig problem is fixed now. Thanks!
Using Linux 64 bits