Is Frame Rate a Primary Workload for GPUs?

I think there's a couple drivers the idea that you can sort of algorithmically imbue this algorithm with an awareness of the hardware is something which is hard to sort of learn at scale. The other thing too is just raw experimentation. I mean if NAS can produce a thousand candidate models in some unit time how long would it take a data scientist to think up the same thousand experiments right? Yeah so that's kind of a complicated surface. We didn't talk about for example mixed precision right is another kind of technique of like you don't have to just do stuff in one you can mix different layers and different precision and we support that.