That's an interesting idea. It might be a good way to circumnavigate this problem. But there are some drawbacks. Maybe there's a lot of things we could get away with actually needing FP16 accuracy (like iterative methods can sometimes do this, especially in ML) but call FP32 but there's plenty of times where FP32 matters. So I guess it is highly dependent upon those issues and where you can get away with them. But further, how do you enforce that? I think it is interesting though.