And can someone remind me the reason no one uses the 3dfx SLI method?
I assume you mean real Scan-Line Interleaving: even rows processing on one chip, odd rows on the other. I think it's not used because it breaks many of the techniques current GPUs use for memory efficiency.
Post-rasterization, rendering is inherently 2D. Pixels are processed in quads, quads are usually part of larger 2D tiles. Textures and rendertargets are stored in memory "swizzled" or "blocked" or "tiled" (pick your favorite term): a contiguous sequence of bytes in memory stores a 2D region of pixels, not just a series of pixels in a single row. Most color and depth compression schemes are also based on 2D tiles.
With interleaving every other line it's hard not to lose efficiency. If your "quad" now contains pixels two lines apart, you're going to get worse texture cache behavior unless the cache and the texture layout in memory are redesigned to account for the interleave. (More fundamentally, your derivatives are going to be even worse than they already are.)
AMD's supertile method is a form of interleaving that makes much more sense in a 2D world than fine-grained interleaving along one dimension only. But load balancing interleaved schemes is harder than just moving a split line up or down the screen. With a split line, if one GPU takes longer to do its part than the other, decrease the size that GPU is responsible for -- that part must have more graphics load. With supertiles, how do you know which tiles are "heavy" and which ones "light" so you can redistribute them?
And of course, none of these interleaving schemes help you scale your geometry throughput, which is part of the reason both Nvidia and AMD seem to prefer interleaving on alternate frames these days.
