To answer the question: I would be quite confident with a 4xRG Supersampling algorithm, that would deliver equal or almost equal to 4xLevel anisotropic filtering, if it wouldn't put such an ungodly stress on fillrates/bandwidth.
Today's known supersampling algorithms filter textures, but only up to (or close to) 2xLevel anisotropic.
Since for the above accelerators will run into obvious bottlenecks, I'd then like to see a far more advanced hybrid MS/SS sampling sollution to be combined with a truly adaptive sampling anisotropic algorithm, where the number of samples (between 1x and 16x) can vary depending on the surface.