In AFR mode each frame is separate, so there should be no increase of traffic at all. DX11 tessellation does not reuse any data or store any additional data to GPU memory. Everything is generated on fly (including tess factors for all edges and for inner triangle).
DX11 tessellator is designed so that it doesn't need to waste GPU memory bandwidth storing the tessellated vertices first to memory and read them back later (like some earlier designs did). It's basically a layer between the vertex shader and the pixel shader. Vertices are the most coarse data, tessellated vertices are more detailed and pixels are the most detailed. Tessellator actually saves memory bandwidth as the vertex data can be more coarse (less vertex buffer reads), and the whole tessellation step reads from the post transform vertex cache and outputs to a similar cache.
The tessellation shader needs to impement 3 methods that the tessellator calls (control point method, patch method and vertex calculation method). This design clearly tells that the whole patch (usually a single untessellated triangle) is not tessellated first completely and then read, but instead the system calls the methods when it needs more data to triangulate. The tessellated mesh should always be more efficient to render compared to a equal static mesh. So it's pointless to store the last frame data, as it should be more efficient to generate the new data (than read & write huge amount of tessellated data from/to the GPU memory).
