Note that this won't work with reasonably performant CNNs. Passing an image batch through a large-ish ResNet takes half a second on our GPUs, several minutes at full load on CPU. This makes training infeasible, and most models small enough to work on CPU are so far from state-of-the-art that you can't do any worthwhile computer vision research with them.
Yes, but note on the other hand that simpler infrastructures such as one-digit-wide-GB GPUs you could buy and install on your workstation could be similarly frustrating, because you may easily encounter their limits (as in, "I got this semi-specialized equipment and I cannot get an output above 1024x768?!").
So, while one is learning, the case could be for being conservative and work directly on available tools, which will be revealing on some scalability requirements, also optimistically: you do not need a full lab to do (reasonable) linear regression, nor to train networks for OCR, largely not to get acquainted with the various techniques in the discipline.
When the needs push, it sometimes will not be just high-end consumer equipment to solve your problem, so on the side of hardware already some practical notion of actual constraints of scale will help orientation. Because you do not need a GPU for most pathfinding (nor for getting a decent grasp of the techniques I am aware of), and when you will want to produce new masterpieces from a Rembrandt "ROM construct"¹ (and much humbler projects) a GPU will not suffice.
(¹reprising the Dixie Flatline module in William Gibson's Neuromancer)
...I am curious, now that I know about Fabrice Bellard's LibNC (bellard.org/libnc), if that «image batch through a large-ish ResNet» would be faster using this library - which can work on both CPU and CUDA...
