True. But nummerical stability is much more important in long running simulations like weather forecast, than in deep neural network training.

There is a reason they are often benchmarked with single or even half precision.