On average, what would the fp16/32 ratio be for graphically demanding games? At fp32, the TX1 is capable of 0.5TF. So if say 1/3rd of instructions runs on fp16, that means an equivalent of 0.66TF, for instance, right?
Yes, for such a hypothetical split that would be correct. Of course it all comes down to the split. The assessment of that could be tricky, but let's try to hypothesize:
1. Vertex positions generally need maximum precision - 1-unit errors in viewport coords can cause high-visibility high-freq artifacts - shimmerings, etc. So all vertex pos transformations are better left in fp32.
2. Likewise with geometry shaders - while some base computations could be done in fp16, the eventual MVP transform is best left in fp32 - i.e. all geometry-shaded vertices have to eventually pass a fp32 MVP.
3. Everything that ends up as shading though is very error-resilient - starting from normal transformations for lambertian terms, color blends and factorisations - pretty much everything which is not a tex coord for a dependent texture read could go as fp16. And then even some tex coords could go fp16, given a sufficiently-low tex res.
4. An exception to (3) are shadows - PCF shadows, etc, could exhibit crawling artifacts and/or perforations if stored or carried (read: sampled and/or filtered) in under-precision, particularly along the edges.
5. Then come post-effects, where tone-mapping could be a real wild-card - one might or might not get banding artifacts if those were carried in fp16 depending on the paramers of the scene lighting. But let's say that to be on the safe side fp32 would be the preferred format.
Again, these are all off-the-top-of-my-hat musings. One normally takes a pipeline, sits there with a few shaders in mind and experiments.