I think I agree with you with the caveat that it would depend on what they're researching. If they're researching new model architectures, I don't think it mak

I think I agree with you with the caveat that it would depend on what they're researching. If they're researching new model architectures, I don't think it makes sense for them to try to implement the algorithms from scratch in C++/CUDA to do distributed GPU training--why not just use TensorFlow? But if you're researching distributed tensor computation, then that's your bread and butter.