So here's something interesting. I was skimming through Cadence's blog, and particularly their
report on TSMC's most recent tech symposium last month. While reading through their part on the InFO packaging tech, this particular line caught my eye:
Note that they're talking about a singular four-die HBM stack. Assuming we're talking about HBM2 (there's not really any way it could be anything else) this means it's a 4GB stack with bandwidth around 200-256GB/s. This means it's not a GPU, as it's neither enough bandwidth nor capacity for high-end cards, and it's going to be too expensive for more entry-level cards (where 200GB/s can be easily achieved with much cheaper GDDR5).
That leaves SoCs. Xavier is obviously the most likely candidate, although even that would be a bit of a surprise. Not just the bandwidth (I honestly have no idea what kind of bandwidth deep learning requires), but that I'd be surprised that they would be so heavily power constrained as to avoid GDDR5(/X), which could also give them much more capacity for their money.
The other option is NX. HBM2 would, of course, be absolutely insane overkill for a handheld. There's no way Nintendo would next more bandwidth than PS4 Pro, with a system perhaps a tenth as powerful and a vastly more bandwidth efficient GPU architecture. And yet, from Gamecube/Wii's 1T-SRAM to 3DS's FCRAM and ample SRAM, to Wii U's big eDRAM pool, Nintendo clearly loves expensive specialist RAM which is tightly coupled to the CPU and GPU. I would also expect HBM2 to consume a bit too much power for a mobile application, but we're only talking about a single stack, and I suppose it's possible that they could be clocking it down to 100-150GB/s to sit at a more comfortable point in the power curve. Lastly is obviously the expense, but only using a single die and InFO (which avoids the need for a silicon substrate) might just potentially pull this down to less than catastrophic level.
If I were a betting man I'd certainly put my money on Xavier being the chip they're talking about, and Nintendo going with a more typical LPDDR4 solution perhaps with a decent sized shared L3 cache. That said, it would be the first Nintendo device in a long time to go with the "typical solution" for RAM.