Comment by vbezhenar

It's still now clear how small rounding change results with NaN. It would be understandable if some triangle would be off by one pixel on AMD and probably not noticeable.

So far it looks like a CPU bug which produces NaN on some input.