l$.$ You train a very large model, which takes many hours to train.

ii$.$ You train $10$ slightly smaller models and average the results of each. The

entire process takes a smaller amount of time compared to the first

model.

Both these models give you an accuracy of $90\%$ on your training set. You need to

pick an approach to deploy on a single$-$GPU machine that needs to provide

predictions in real$-$time. Quick turn$-$around is of the utmost importance here.

Which approach is more suitable for production, and why?