Or perhaps a blitter capable of rasterizing triangles with affine texture mapping. A bit like PS1. Low silicon requirements as just addition is required. Nothing complicated like multiplication or divides needed. Of course the price is loss of perspective correctness.
Even AGA was a kludge - planning was started once they realised their "real" next gen chipset was going to take way too long.
Bus width went from 16 bit to 32 bit and fast page mode provided another bw doubling.