# How to calculate computational efficiency of Deep Learning Models?

I am trying to make a comparison between two simple 5 layer neural network models. One of the models has 3 frozen layers as I've implemented transfer learning in that architecture. The other is identical to the previous one but is without the frozen weights, i.e., all the weights in this network are learnt in the training process. The training process is also identical (batch-size, epochs, loss function etc.).

I want to try and compare the different metrics that evaluate their performance, not accuracy-wise but efficiency-wise. To do this, I used tensorboard and a library called WeightsAndBiases for seeing the system metrics. In TensorBoard, I used the metric Average Step time, and in WeightsAndBiases, I used the metrics CPU Usage and GPU Usage.

But what I recently came across is that the GPU usage should be as close to 100% as possible (It was a maximum of around 26% in my case). Does that mean using CPU/GPU usages would lead to a false inference (even if the training process was the same for both)? If yes, are there some additional metrics that I can calculate to show which model is more computationally efficient?

Also, if the answer is yes, why is it so? In my opinion, if the training is identical for both the architectures, it would mean that whatever difference in the metrics exists, it would exist because of the computational load of the models.