And this is where development on PS4 Pro will actually greatly begin to benefit the Switch with third parties, as new techniques for taking advantage of 2x FP16 will be developed.
There will be a bit of spillover from PS4 Pro and Scorpio supporting FP16, but I wouldn't expect anything major in the short-term. The impression I've got from reading about PS4 Pro support indicates that it tends to be rather small teams working on them, who would be unlikely to take the time to optimise for something like that, which wouldn't carry over to the other version(s) of the game. I'd expect the bigger middleware engines (e.g. UE4) to do so, and some internal teams like Naughty Dog would probably put quite a bit of work into optimising for PS4 Pro, but I wouldn't expect that much from internal engines by most third parties, as they're likely to focus on methods for improving performance on all platforms, and particularly on the ones with the largest install base (which is still going to be the regular PS4 and XBO for some time).
Anyway, what are your thoughts on the ability for a standard TX1 (like that in the Shield TV) to actually maintain the claimed clock speeds for 8 days straight? Presumably, since the Shield TV does indeed throttle according to MDave to levels below those from this leak, then that would indicate that the SoC being tested with those leaked clock speeds cannot be a standard TX1 right? It would suggest 16nm but I guess it could also suggest an SM or CPU core is disabled.
I just re-read through the updated translation, and it doesn't seem to actually mention testing for 8 days (unless I'm missing it). The only similar point it mentions is "There's no lag whatsoever after running for 2 hours straight"
In any case, it doesn't matter too much whether it's 2 hours or 8 days, if it's stable for one it's likely to be stable for the other.
The question would really come down to cooling. Electrically, we know the TX1 can hit those clocks (otherwise the Shield TV would have shut down immediately in MDave's test, rather than throttling), but it reaches a sufficiently high temperature to throttle in the Shield TV's case and with the Shield TV's cooling system. While we might expect Switch's cooling system to be less effective than Shield TV's, due to the smaller case, that's not necessarily guaranteed. There are a number of aspects of cooling design where, in theory, Nintendo may have improved over Shield TV's cooling system, such as using copper rather than aluminium cooling fins, using higher quality heatpipes, having a better thermal interface between chip and heatpipe, or between heatpipe and cooling fins, using a higher quality and/or higher speed fan, or just designing the system's internals in such a way as to facilitate better airflow.
We can't necessarily guarantee that Nintendo has done any of these, but we also couldn't rule them out. It should be possible to cool a 20nm chip at those clocks in a case the size of Switch, but it would likely require a pretty well designed cooling system.
The two problems I have with the 1.785GHz clock being used for games aren't related to the ability to power or cool the chip, but rather that (a) it's an extremely large jump over a 1Ghz clock which had already been described as final and (b) Nintendo seems intent to keep CPU clocks the same between docked and portable mode for games, and it seems like a very high clock to retain a decent battery life in portable mode (although this of course depends on architecture and manufacturing process).
If you pay close attention you'll notice the side profile of device's side matching the switch side profile to the t, the top bevel included, all underneath some form of a lid/cover placed over the joycon rails - notice the different plastic texture of the lids, vis-a-vis device's chassis.
That's a fair point. I had assumed the side panels were screwed on in place of the rails, but it's interesting that they have that little lip at the top. It seems a little odd that they would feel the need to cover up the rails with anything, as the photo was obviously only ever supposed to be seen by certified developers in the first place.
Thanks for the explanation, looks like my understanding of fp16/fp32 was a bit too limited. If they can actually get that 70% ratio of FP16 that Ubi dev mentioned, then that would mean quite a significant boost (54% if
my math is correct).
Yeah, seems about right. It would obviously vary quite a bit from one game to the next, probably with better utilisation with Nintendo's internal engines. It's also not necessarily guaranteed that pure computational throughput is going to be the bottleneck in any specific scenario, but in cases where the game is tightly optimised for the architecture (i.e. first party games), it could potentially make a decent bit of a difference.
The points I bolded from the translation quote perfectly describe a home console IMO.
I dont believe it's a Dev unit. Why would the switch Dev unit be ps4pro power levels when the switch is multitudes lower in power? I hope we get more leaks about this from someone. Can't be at Nintendo's mercy to release info.
The unit is described as an "enhancer" which attaches to the Switch, and is used in addition to the Switch's SoC and RAM. This would point towards it being some kind of add-on rather than a stand-alone console.
The argument that it's a dev unit isn't that it's a dev unit for the regular Switch (as you say there's no reason for an extra GPU in that case), but rather that it's a dev unit for some kind of future hardware, possibly an add-on for the Switch which is still in development.
It could easily be a different version, but I still think he's talking about something resembling the picture.
The picture most likely has either a standard or modified Switch motherboard, with a second board attached to the back of it via some kind of PCI bridge, no?
Yeah, I wouldn't be surprised if it did look somewhat like the unit that was pictured, I just don't think it was that one in particular. For one, I'd expect it to be somewhat larger if it's cooling both Switch's SoC and a separate 200mm² chip.
I completely agree with this. It doesn't make any sense for Nintendo to pursue a 4k dock with the Switch, at least as far as I can tell from Nintendo's market. Maybe they have market research which says otherwise but I doubt it.
The question remains though, what is this leak about? It's very hard to just toss out any portion of this leak considering how much info he got right, so what is he describing? Maybe a prototype unit for a standalone console? Maybe a VR prototype?
As Cuburt said above, the SCD patent had much more interesting potential as a local cloud device, and Cuburt brought up an excellent thought in that purchasing an SCD could grant that user a free online subscription, as the SCD would/could wind up strengthening online infrastructure. Could the kind of die (20x20mm) described by the leaker be explicitly purposed as a kind of "local cloud server"?
Something like that seems like a much more beneficial (and "Nintendo") avenue for Nintendo to go than a 4k dock.
I agree that it wouldn't seem to make much sense from a business perspective, but at the same time the other possibilities seem to make even less sense to me. I would narrow down the possibilities into two groups as follows:
- Possibility 1: The 200mm² chip is an existing "off the shelf" chip
By this I mean it's not a chip designed custom for Nintendo. It may be a stand-in for a custom chip, or they may for whatever reason decide to actually use an off the shelf chip for the final product.
In this case, it would seem reasonable to restrict the possibilities to the chips Nvidia produce or have in production. It's certainly technically possible for them to use a chip from another source, but it would be extremely bizarre, for example, for them to use an AMD chip in the add-on to their Nvidia-powered Switch. Nvidia make GPUs and SoCs, and on the GPU front they have one model which matches the described dimensions almost exactly, which is the GP106. Their SoC lineup doesn't have a matching model, though. The TX1 and Parker are both smaller than the described chip, and Xavier is expected to be bigger (it also isn't due to sample until late this year).
Out of the plausible candidates for a pre-existing chip that Nintendo may be using, the GP106 would seem like the most likely option. It would also make sense as a stand-in for an existing chip, either a dedicated GPU or an SoC of some sort. The timing would, then, also make some degree of sense. A device being developed with an intended launch of late 2018 or thereabouts would expect to have its first run of dev kits go out around now, with off the shelf hardware used in place of custom chip(s) still being finalised. It would also explain why we haven't heard anything about it from developers yet, as if it's early enough to still be using stand-in hardware it may not have made it to many (or any) third-parties yet.
- Possibility 2: The 200mm² chip is a custom chip designed specifically for Nintendo
In a theoretical sense, this chip could be pretty much anything. It could be an SoC, a GPU, a CPU, or a fixed-function chip of any kind. The potential applications are quite wide.
The issue with this possibility, though, is the timing. For Nintendo to have 2000 samples of a custom chip to be ready to go out to developers, they would have to be very far along in development. Keep in mind that final Switch SoCs reportedly only hit third parties a few months prior, and it would suggest a launch timetable perhaps 6 months or less after Switch itself. This would seem very unlikely to me, firstly from Nintendo's point of view to release a significant "upgrade" to Switch after it's barely hit shelves, but also because we haven't heard anything about it until now. If this had been in development concurrently with Switch for 1-2 years, we would have expected at least some kind of hint of it in one of the numerous leaks we've had on Switch over the past 6 months or so. We haven't, though, which would indicate that it either hasn't hit third parties yet, or it has hit very few of them, and in either case that is evidence of a device early in its development cycle.
Regarding the concept of a "local cloud server", you'd have to be a bit more specific as to what that entails. From the simplest of implementations being a wireless storage device (a NAS, effectively), it certainly wouldn't need a chip that large powering it. It could provide meaningful computational capacity (i.e. CPUs and/or GPU) over a wireless connection, but if it did so I would imagine it would be in addition to a hard-wired connection rather than as a replacement for, given the significantly enhanced capabilities a hard-wired connection would provide. The use of a PCIe bridge with the Switch (or a version thereof) in the dev kit would suggest to me a device that is intended to physically connect to Switch, at least in some circumstances.
Prior to all this I would have said that the most sensible way for Nintendo to pursue something like this would be to just release a traditional home console device (let's call it Switch Home, for the sake of argument). This would have a wider potential audience than a dock, and given Switch's specs it could likely play most Switch games at 4K while keeping the cost low. It could potentially also play third party games at 1080p if Nintendo allowed developers to make game supporting the Switch Home, but not the regular Switch. It would also be relatively low-risk, as all of Nintendo's development efforts would still benefit owners of other Switch devices.
I still think that seems like the most sensible way to go, but the evidence seems to suggest that the "enhancer" is something different. It's certainly possible that they're using the GP106 as a stand-in for a new custom SoC, with a few A72 cores and perhaps a GTX1050-level GPU. In that case, it may make sense to use both the Switch SoC and the GP106 combined together on the same board to stand-in for the new chip. The problem is, though, that this isn't a standalone box with both chips on one board, it's a device which attaches to Switch via a connector, and when combined they still have a screen (and can presumably be detached). It wouldn't seem to make much sense to design a dev kit in such a way if the device doesn't make any use of the regular Switch's functionality.
I suppose one option which is technically possible is that it's a stationary home console that
also works as a dock for Switch. That is, when you buy it it comes with a pro controller (or possibly a pair of joycons) and can be used on its own without a Switch plugged in. If, however, you chose to plug a Switch into it you could use the same docking and undocking functionality as normal. It would be somewhat complicated from both a hardware and software design point of view (even more so than a GPU dock), but not necessarily impossible.
so ... how would undocking in the middle of a game work?
wouldnt the SCD have all the GPU memory (assuming 4k would need more than the 4Gig that in the base console)? ... with the current configuration, all that data is in the console so undocking doesnt cause any disconnect between CPU and GPU/Memeory
It's certainly possible, and laptops have been doing a similar thing for quite a while with automatic switching between a dedicated GPU (with its own RAM) and integrated GPU (which shares system RAM). Microsoft's Surface Book even allows users to physically separate the keyboard section (which includes a dGPU) and the system will revert to the integrated GPU automatically.
It is a little more complex in Switch, with a lower-level API and software specifically tailored to the hardware, but not impossibly so. It would likely be a matter of how much of a delay they'd be willing to allow for after you dock or undock the Switch. The Switch's memory pool would have to store all data relating to the game state, with the GPU's memory pool storing assets (i.e. textures, models, etc.) and render targets. This would pretty much be the expected division of data anyway, though, so would be unlikely to be an enormous burden to developers.
The decision then is whether at un-docking time you just re-load all assets from the game card or flash memory, which would be relatively straightforward but slow (effectively you're just loading the game again at the precise point you left it) or you could store lower-detail assets in Switch's memory pool to be prepared for undocking. This would reduce the delay after undocking, but would use up memory which then couldn't be used for other things, and would also be more complex for developers to implement.