Allocating more bits for each parameter increases precision, by definition. But that doesn’t come for free.* So it is useful to optimize network performance for a given number of total parameter bytes.

I haven’t done a recent literature review, but my hand-wavy guessplanation is that a NN (as a whole) can adapt to relatively low precision parameters. Up to a point.

* In general. Given actual hardware designs, there are places where you have slack in the system. So adding some extra parameters, e.g. to fully utilize a GPU’s core’s threads (e.g. 32), might actually cost you nothing.