There is not even such an improvement in power efficiency between 28nm Maxwell and 16nm FinFet Pascal, and 20nm at low frequencies (Like the Switch uses) should have the power efficiency advantages of a die shrink compared to 28nm. There is a reason Nvidia chose to implement it in the Tegra X1 but not in Maxwell parts that ran at high frequencies like PC gaming cards.
My opinion is, the upgrade will be completely silent. We'll need 7nm in late 2020 or so for a leap over the current Nintendo Switch.
About a possible home console, a 16nm version of the current chip clocked 80% higher and with 8GB of RAM and twice the bandwidth would be sitting in XBOX One territory... But I doubt something like that will happen.
I meant 50% more as an upper bound than anything else, probably should have been more clear. My point was just that, even if they increased clocks by 50%, it still wouldn't be a big enough jump in performance to warrant a new Switch Pro, or whatever people would want to call it.
I agree that #1 seems like the likeliest thing here. I doubt it's even "based on TX2" as much as it's just a die shrunk TX1, at least going by the SoC number.
I'm curious what other changes they may make to the new Switch though, besides fixing the security issue. Like, a Thunderbolt ready USB-C port? Maybe double the RAM bandwidth?
Thunderbolt is both very expensive, and not much use on a new Switch (particularly if it wasn't on the original model). It's also only ever been used on Intel-powered products, although Intel have said they'll be opening it up, I don't know if there's been much progress on that.
Regarding memory bandwidth, there's not a whole lot of point increasing it if you're not increasing the other specs along the way. You'd get a bit of a benefit in bandwidth-limited scenes in a handful of games, but it would be of sporadic benefit and I can't imagine most players would notice the difference.
I'm pretty sure it's also possible to get 8GB of LPDDR4 on the same 64-bit bus as the current Switch, by the way (not that I think they're going to release a Switch successor with 8GB RAM any time soon, but FYI).
I think it's fair to say that memory bandwidth is one of the likely culprits for games having lowe than expected resolution. Certainly problematic with alpha effects regardless of compression.
Lower than what expected resolution? People seem to be expecting Switch to render everything at a constant 1080p, but XBO doesn't come close to that with most third party games, and it has a larger SoC operating at a far higher power draw.
The only problem I see with the theory 1 is why double the ram? The reason that does make most sense to me is they are using a 128bit bus width, which would require 8gb to get up to double bandwidth? And if they are doubling the bandwidth and the amount, that would indicate a more powerful model.
The logic would be that 1 and 2 would both use 4GB of RAM, as with the current Switch. The 8GB would be just for the home console.