The claim about the SNES's memory access is rather loaded. At 3.58 MHz, it does complete a memory access at about 86-88% faster (280 ns vs 522 ns), but that's for an 8-bit read vs a 16-bit read on the MD. (so the MD takes ~1.87x times as long, but can read double the data in that time) And then there's the bigger issue of the SNES actually runs at 2.68 MHz for all games in RAM (and all early games in ROM as well), so that's only 373 ns accesses and the MD only takes 1.39x as long per access then. (but still can do 2 bytes in that time when the SNES is handling 1 byte)
And that's just memory accesses, not actually addressing performance benchmarks for running code. (the SNES executes most instructions much faster, but the simpler instruction set means more instructions needed for some operations the 68k does with fewer, slower/more powerful instructions -to which point it depends on the type of game, but the SNES is so slow that it still will only come close to even in better cases, more so with later 3.58 MHz games)
The PCE's CPU accesses ROM and RAM at 140 ns, so 273% faster than the MD (MD takes 3.73x as long to complete an access), but still 8 vs 16 bit meaning the PCE has a nominal ~87% bandwidth advantage over the MD's 68k. (double that for 8-bit specific operations -the same cases where the SNES would have a moderate advantage over MD bandwidth)
Then there's the actual computational performance again, still cases where the 68k could be preferable, but a much wider margin for the PCE's CPU to match to beat the 68k in other cases. (any case where the SNES would be closer to even with the MD, the PCE should beat both by a fair margin -some of the biggest exceptions where the MD would be better could be for 3D games, though the MD has a bigger advantage there with packed pixels vs planar on the PCE, so it wouldn't be an even comparison in any case)