So after watching the Digital Foundry video and skimming through the dozens of pages added in the last couple of days, I have some thoughts, which I'll try to cover in the most coherent way I can manage.
First off, to reiterate the two pieces of info from the DF video:
- There was as some point a mid-gen refresh of the Switch planned, which was cancelled (this was backed by conversations with developers)
- A couple of the DF team think a 2023 release for a next-gen Switch is unlikely (this was presented as opinion)
They don't specify what kind of hardware was due to be in this mid-gen refresh, when it was planned, or how far along it got before being cancelled. As I see it, there are three possibilities for the hardware side of things:
Mid-Gen Possibility 1: Overclocked Mariko
This seems plausible to me. We know Mariko was available since 2019, and we know it would have been able to clock higher than the original TX1 and supported faster RAM. A timeframe of 2019-2021 would make sense for a device like this, and that overlaps with Bloomberg reports that a mid-gen refresh was planned for 2021. On the negative side, an overclocked Mariko wouldn't have been a huge bump in performance. Possibly a 50% boost in clocks at the highest end of expectation, maybe with additional RAM too. It would be a noticeable improvement, but it definitely wouldn't have been capable of pushing Switch games to 4K (as Bloomberg claimed from the new device).
Mid-Gen Possibility 2: A Cancelled Chip Between Mariko and Drake
I would consider this unlikely. It's definitely possible that Nintendo and Nvidia discussed various hardware options before Drake, and some of them may have made their way into the planning stages. However, this report comes from speaking with developers, and for developers to start work on software for the new device they need an SDK which supports that new hardware in some way. The Nvidia hack from earlier this year, though, included the source code to the graphics SDK that Switch developers use, and there are apparently no references to any hardware between TX1 and Drake in there. If an intermediate chip was developed to the point of third party developers actually working on games for it, then you'd expect vestigial references to it to be littered all over the NVN source code. The lack of these makes an intermediate chip very unlikely in my view.
Mid-Gen Possibility 3: Drake
I also consider this unlikely. As I've said since the details revealed in the Nvidia hack in March, Drake is simply overkill for a mid-gen refresh. If you want hardware to run Switch games at higher resolutions like 4K, there are two options. First, just use a bigger Maxwell/Pascal GPU (say 6-8 SMs) and natively render at 4K. Second, use a smaller Turing/Ampere GPU (say 4-6 SMs) and render at ~1080p, then use DLSS to get to 4K. The former is a bit more "brute force", but keeps development simple, as the architecture is the same. The latter is probably more efficient, but requires extra work from developers to leverage. Drake is not only bigger than it needs to be to brute-force native 4K on Switch games, but it also has a new architecture with tensor cores for DLSS, full ray-tracing support, etc. I can't imagine a world where Nintendo went into a design process for a mid-gen refresh and ended up with Drake.
As a sort of possibility 3.5, I've also seen people propose that Drake has changed over time, with an earlier version of it being intended for a mid-gen refresh, then it was upgraded to a larger chip for a Switch 2. I don't personally think this is likely for the same reason as possibility 2 above; there would have been some vestigial references in the Nvidia hack to the older version of the chip. However, as far as I'm aware, there's no evidence in there that T239 was ever anything else than it currently is.
For me an overclocked Mariko seems the best guess, although I'll admit it would have been a somewhat meagre mid-gen upgrade. It also wouldn't have really aligned with Bloomberg's claims that Nintendo's plans were for a 4K device, or that they would use DLSS. It could be that his sources were referring to two different devices, one mid-gen upgrade which was cancelled, and the other being the Drake-powered Switch 2, and they got conflated into a single model.
In terms of timing, although the comments on the likelihood of a 2023 launch was only speculation on their part, it does make a H1 2023 (ie alongside ToTK) launch much less likely in my eyes, as if the new hardware were 6 months away I would assume they'd have heard of it. I still wouldn't fully rule out H2 2023, though.
I think it's worth coming back to what other info we have, though. Between the Nvidia hack, and a variety of L4T (Linux For Tegra) commits from Nvidia, we know that work started on Drake/T239 in late 2019/early 2020, and that it's a chip which is overkill for a mid-gen upgrade, but well-specced for a Switch 2. The less-capable tensor cores compared to Orin, along with the lack of DLA/PVA/etc., point to it being poorly suited for Nvidia's other SoC use-cases (ie automotive and Jetson), and with the file decompression engine it looks very much like T239 has been designed specifically for Nintendo. The presence of T239-specific code in L4T suggests that it will also be used on non-Nintendo products, similarly to how Mariko was also used on the Shield TV.
There are two more recent indications that some degree of manufacturing has started on Drake. Firstly, there have been a couple of T239-specific commits upstreamed to the Linux kernel (ie they're in standard Linux now, not just L4T). There's no iron-clad rule on this, but generally companies don't upstream this kind of thing until manufacturing has started and they're preparing to release the new hardware. Secondly,
@oldpuck found a reference to T239 PCIe timings in a L4T commit. I can't remember the specifics off the top of my head, but it's the kind of thing which only makes sense if Nvidia have actual silicon on hand to work with. Neither of these are enough to prove that the chip is in full scale manufacturing, but they suggest that at the very least engineering samples are available, and full-scale manufacturing typically isn't too far away at that point.
To me, the most likely course of events for a chip like Drake is that it would be manufactured firstly for Nintendo, and then Nintendo would allow Nvidia to use it in other (probably non-competing) products after Nintendo's shipments are fulfilled. We saw this on Mariko, for example, where a new Shield TV showed up shortly after the Mariko Switch models. Nintendo accounts for 99+% of Nvidia's consumer SoC business, so manufacturing an SoC like this without Nintendo doesn't make sense for Nvidia. Meanwhile, if Nintendo get to dictate the design of the chip, and then can get a better deal by allowing Nvidia to also use the chip in some other products, they might as well do so.
The evidence that Drake is entering manufacturing is therefore important, as I don't believe this would happen without Nintendo's go-ahead. Had Nintendo decided to cancel their Drake-based model before manufacturing started, or delayed it by a year or more, then I can't see Nvidia going ahead with manufacturing. They just don't have a big enough market for a chip like this without Nintendo on board. It's technically possible that Nintendo could have cancelled or delayed the device
after manufacturing had started, but that would be the worst possible time to do so, costing them a vast amount of money.
To me this still points to a 2023 launch being on the cards, although more likely late 2023 given DF's comments. Either that or Nvidia have decided to go ahead with manufacturing on T239 on their own without Nintendo, which would require a much bigger product than the Shield TV to justify. Maybe they have such a product (I'm genuinely struggling to think of one), but I personally don't see it happening without Nintendo. Ergo, I still see a 2023 release as on the table, even if DF don't.
I won't comment on NateDrake's posts yet, as I'll have to go through and read them again, but I may wait for him to post a video on it so that he can present things more cohesively.