*cracks his layman fingers*
With VRS you can decouple the Rasterization-Rate (How fine your pixelgrid is) from the Shading-Rate (shading/coloring the pixelgrid).
That's the difference to how we do it now where we scale both at the same time, either up or down:
Now as far as I understand it you wouldn't get more sub-pixel polygons with VRS since the amount of raw pixels per polygon is still the same but you can save shading costs when you let the pixel shader run for multiple raw pixels at once instead of each individually.
Overshading with small triangles is an issue, potentially with VRS you could save a lot but the question is at what visual cost?
If you need precise shading for small details you might not want to reduce the shading rate there.
The classic rasterization/shading pipeline is quite complex from a layman's perspective.
You have fixed function hardware for geometry generation but programmable shaders are also involved in the process.
Triangle culling can be done at different points with various precision and efficiency.
There is fixed function HW which does the rasterization of the scene, also with varying granularity, pixel shaders are actually doing the shading processing in pixel-quads (2x2) and not individual pixels (with VRS it get's more complex to understand) since that way you get derivatives for texture filtering, you have HW for Coverage/Z/Stencil-Testing of the pixel-quad-packages, you have to follow the submission order to guarentee correct results, some stuff in the pipeline is getting serialized because of it.
I'm not in the know if and how much VRS and Mesh Shaders ultimately could help to improve the efficiency of the whole pipeline if micro polys are the main issue.
I would guess that VRS won't help much here but Mesh Shaders and feeding the machine only with visible triangles and with knowing the triangle topology you could shade meshes with far less unnecessary computations.
But performance is of course a general resource.
If you do stuff more efficiently you can also spent more on other areas or afford more of sub-optimal rendering for better visuals.
I know my posting boils down to basically this:
But there is google and I know that there is at least one (older) patent from AMD about improving the performance in the case of micro polygon rasterization, if you really want to know about the bottlenecks and potential improvements then there might be some answers.
I have more or less a list of 20 patents I want to read but no the time for all of it.