I agree that, in theory, a pool of fast embedded memory plus a larger, slower main memory pool seems well suited to a game console, but I'm becoming less convinced that it's a smart choice in practice, when faced with a limited budget. Of course it's near impossible to say whether a, say, 128-bit GDDR5 pool would have been cheaper for Nintendo than the eDRAM + DDR3, but we do have a near-perfect case study of the two approaches in the PS4 and XBO.
Both had access to identical CPU/GPU architectures and a very similar BoM (judging by sales price once MS dropped Kinect). Pretty much the only meaningful high-level distinction between the two designs was MS's teams's decision to combine embedded memory with DDR3, and Sony's decision to go with a single GDDR5 pool. The results are quite obvious; the single pool was the better decision. That's not to say that MS's embedded memory approach didn't have its advantages, but they were obviously heavily outweighed by the extent to which they had to cut back on GPU logic in order to accommodate the memory pool on-die.
Granted, XBO's 32MB may be less than ideal for its target resolutions, and they used SRAM rather than Wii U's (presumably) cheaper eDRAM, but in the absence of other evidence I'd err on the side of a single fast memory pool being the better approach, either to maximise performance at a given cost, or minimise cost at a given level of performance.
This, of course, excludes the possibility of a tile-based GPU like Nintendo has moved to with Switch, which would obviously have a different set of trade-offs.