I'm actually a lot less sceptical of Nintendo revealing detailed specs of Switch since they publicly talked about their new NVN API. This is the kind of thing Nintendo
never would have talked about before, and it may be a signal that they're changing their attitude towards how open they are with the technical details of their hardware. Nintendo's under new management now, their hardware design team have changed and the way they're presenting Switch seems to be changing too. They could also (with some degree of legitimacy) claim that it's powered by the most graphically capable mobile SoC around, so it may be worth getting Jen-Hsun Huang up on stage during the presentation to do his thing.
Does having higher than expected specs honestly matter that much? Even at a max of 756 GFLOPS in FP32, a memory bottleneck and the fact that this thing most likely won't be running at full power in portable mode automatically rules out most modern third party AAA titles. The way I see it, Nintendo is probably aiming to get the essential third party games for the mainstream market (sports titles, Minecraft, maybe some shooters and Rocket League), while offering their first party titles and some Japanese games as the system's AAA. I anticipate a lot of remastered games from last gen (or last last seeing as this is 9th now I guess) since those are popular on XBO/PS4, though Switch's versions can be marketed with having the advantage of portability. Indies will presumably pad out the rest.
I'd personally be fine with Wii U quality visuals with better texture filtering at 1080p, or graphics a little below Xbox One running at 900p. The architecture and APIs are a lot more modern than Nintendo's last few systems, and with the most pessimistic estimates pointing at Switch still being twice as strong as the Wii U in raw power, I'm sure my expectations are realistic.
I don't think we have any reason to believe at this stage that it is memory bottlenecked, but even with that I don't think either that or the raw GPU power would "automatically rule out most modern third party AAA titles". The existence or lack thereof of western third party games will come almost entirely down to whether or not there's a business case for the port.
Back in 2012, when Ubisoft wanted to bring Assassin's Creed to Vita, porting the main console titles wasn't an option, so they had to build an entirely new game from scratch in order to cater to that audience. The game (and the system, for that matter) didn't sell amazingly, but Ubisoft reported that they were actually quite happy with the game's sales over the holiday period. It never received a follow-up, though, as the niche audience just couldn't justify the expense of building full-scale games solely for one platform.
Even if Ubisoft have to absolutely butcher the next Assassin's Creed game to get it running on Switch (and I'm under no illusions here, I fully expect that any AAA third party games would be running at 720p with reduced assets and effects in docked mode and at sub-HD in portable mode), the cost of doing so would still be orders of magnitude smaller than the cost of making an entirely new game from scratch. That niche audience that wasn't enough to justify new games on Vita may well be enough to justify ports on Switch.
This is the big difference between Switch and any other attempt in the last ~20 years to offer "console gaming on the go". There may not be a larger group of people looking for this experience than there were with PSP or Vita, but the cost of serving those people is far lower, so it's entirely possible that Nintendo could provide a niche, but profitable, platform for third parties.
NTD has a bunch of former Tegra engineers these days, so it's likely they knew Nvidia's roadmap for quite some time. We'll see. But Nvidia seems to work on a new custom ARM core they plan to use in Xavier instead of adopting A72s.
They haven't actually commented on whether this is a Denver successor or not (to my knowledge), but given that Denver was designed for a ARM/x86 dual-use that's never going to happen, it would make sense for them to switch to a true ARMv8 core. I'd be very surprised if we did see it, though, as Xavier isn't due for another year after Switch.
Xavier also looks to be aiming for 20W of power draw, even on 16nm. Nvidia seems to want to replicate the performance of the Drive PX2 on one chip, which is great for automotive functions, but clearly not aimed at mobile. I suppose it would be very Nintendo to use custom ARM cores, but Nvidia made it sound like they were pretty far away.
A57 should be able to hit fairly high clocks consistently on 16nm both mobile and docked, but A72 would be much smaller on die and faster to boot. Hard to get a read on where Nintendo's sensibilities would take them in this case.
Xavier is targeting 20W, but that includes a relatively wide and high-clocked GPU, and their 8 CPU cores are likely to be clocked quite high too. It's conceivable that 2 or 4 of them could be clocked more modestly in Switch's thermal envelope (although at this stage I'm really just speculating on a core we know absolutely nothing about).
Which would be cheaper? That would probably be your answer. Kimishima mentioned being profitable out the gate and sensitive to consumer price expectations.
The A72 is smaller (therefore cheaper) as well as giving higher performance at a lower power draw, so there's no real reason to use A57s unless the design was locked down extremely early (which seems to have been the case with Parker). The A73 is also smaller again, and draws even less power (although performance increases seem fairly small going by Kirin 960 benchmarks), but I doubt it was ready in time for use in Switch.
I think it's gonna be tough to get super specific benches between them when most 53/57s are at 28nm and the 72 is for 16nm. So you'll get benefits purely from the process shrink. ARM claims the 72 is a better design even at the same node and speed.
If you want to at least see some real hardware differences, the Nexus 6P used a Snapdragon 810 (4x53 for power, 4x57 for performance) and the new Pixel uses a Snapdragon 821 (2x72 for power, 2x72 for performance). The new Pixel's bench quite a bit higher than 6P's.
The Snapdragon 820 and 821 use Qualcomm's Kyro cores, not A72s. The Kirin 950 and 955 use A72s, and would be a better comparison point to the 810 (although they're on a more power-efficient 16FF+ manufacturing process). You'll also find A72s on the Snapdragon 650 and 652, although in this case they're manufactured at 28nm, so again it's difficult to make direct comparisons.
I also don't know if it makes much sense for Nintendo to go for a split design like you see. With gaming, you want performance at a certain threshold constantly. That doesn't mean you use it all, but it needs to be available. It's why previous handhelds have used chips clocked way below their maximums. That way they can run at that clock indefinitely (battery not withstanding) with [x] amount of performance.
If NV+N went with A72s, I'd say they'd just go with two of them. More likely, they'd go with four A57s as NV has experience with those specific cores (as their TX1 kits all include the same setup).
I don't think Nvidia's experience with A57s has any real bearing on things. Stock ARM cores are probably one of the most straightforward things to integrate in any SoC, and this is going to be comfortably the highest selling Tegra chip ever, so they're not going to skimp on the design costs.
I do think 2 A72s might be plausible, though. I had previously suggested that a 2:4:2 configuration, with 2 A72 cores and 4 A53 cores for games, and then 2 A35 cores for the OS, might be a sensible direction to take. They'd be able to clock the A72s higher than would be feasible for a quad-core cluster, giving good performance in single-threaded or latency-critical tasks when necessary, and between the six gaming cores there'd be a good amount of performance for multi-threaded tasks at a low power draw. The two A35s would pretty much sip power while performing OS duties (and could be dynamically clocked independent of the gaming cores).
I think it's quite likely that we e.g. still won't know the portable CPU and GPU clock after the "spec" reveal. Even after the release perhaps.
I'd completely agree with this. If they give us numbers, they'll obviously give us the highest possible numbers. We might get leaked clock speeds for handheld mode, or otherwise be able to make somewhat reasonable deductions from the differences in resolution and/or framerate between docked and handheld modes.