Currently the no-GPU switch prevents the GPU from being used at all, even when available. This has the advantage of giving better quality feature extraction outputs, but greatly increases time taken to complete a task due to dense reconstruction not using GPU acceleration.
For many users, using ULTRA feature extraction on full size images is one way to gain the advantage of GPU acceleration for dense reconstruction, and obtaining best quality feature extraction, as the images will not fit in GPU RAM. However, that is not always desirable when images are feature rich, often with half a million or more features detected, high quality feature extraction would be the preferred option.
This thread, where Shiva does extensive analysis/comparison of GPU vs CPU during processing is very informative:
It’s also where I first mention what I’d like to see, ie this feature request 
The race condition I mention in the linked thread sometimes forces me to use no_GPU, at a significant time penalty, if the no_GPU switch only applied to feature extraction it would eliminate that particular race condition error which stops a task from completing.
If you want to use ULTRA feature extraction, then surely it is because you want the best feature extraction possible- which is what you get with CPU, GPU outputs being faster, but not so good quality wise. Quality can be very important for difficult datasets, but in gaining the best quality, you don’t want the time penalty due to not using GPU acceleration for dense reconstruction.
I tried to add GPU as a tag, but apparently it is not allowed in this category.
Should I add a poll to see what other users think?