Well, there are multiple reasons, and the most important one is that T4 is designed for the neural network inference - i.e. for making predictions on the already trained neural networks (especially those that are already quantized and compressed). While...