IMO it's more cost effective and stable just to buy a Ryzen 2400G system (the same cost as a good eGPU case), put in Linux and use it for training at a full speed. TB3 has 4x lower throughput than PCIe, so if you need to transfer a lot of data (any meaningful non-toy model these days), it's better to stick with internal GPU. Not mentioning Linux is 1st class/preferred platform for many frameworks like TensorFlow etc.