Depends on the tools. RedShift for instance told us right after 10xx Nvidia cards came out that if we have lots of vertices (> 1 or 10 million - I don't remember exactly), you benefit from more RAM so get a Titan X. If you want to render anything with less, get more GPUs since you can't share RAM anyways, but the number of cores will help you.
Seemed more of a case that no one wants to actually code for OpenCL on AMD...