Why not just make those external? Makes things easier, especially with cooling. Too bad Apple seems to hate Nvidia. Dunno if it's only where I am, but everyone I know doing ML research uses CUDA exclusively, so AMD is useless. Also too bad the open source libs like PyTorch and Tensorflow are not totally portable and only work without hassle in Linux.
AMD zealots will bombast this and say it's because of price to performance ration, but it has always been speculated that AMD sells it's chips to high profile companies, like Apple, at below production costs. This will allow AMD to seem more competitive professionally than they actually are. If you just go by fastest and best, NVIDIA has been leading for a few years, and now with Pascal and Volta, there is is no question of their dominance. CUDA is closed and proprietary, but NVIDIA architecture it is still the fastest out there.. If Apple doesn't give the ability to choose GPU on the new MacPro, the line is not Pro. eGPU doesn't count..