Now that at least the anger levels have come down a notch, I'll weigh in. Though, I'm waiting for Nate's podcast, or John L to follow up, or some developer to come forward before I speculate too wildly.
I can totally believe the "refresh" was not an overclocked Mariko. Mochizuki said that software using 4k DLSS based dev kits was targeting 2022-2023. It seems like the cancellation of that hardware is the cancelled hardware that Nate and John are referring to. DLSS isn't viable on a Mariko. It seems likely that developers giving Mochi dates were the same ones giving him specs on devkits.
We know Nvidia altered their Orin schedule, moving Nano Next to after AGX launch, but originally having it before. Orin Nano is a floorswept AGX. It seems odd that floorswept hardware would change schedule without base Orin moving. It seems reasonable that Nano Next was, at one point, a different chip than the current Orin Nano.
Or perhaps Orin Nano was expected earlier in the year, as production samples failed to hit yields. Orin Nano 4GB is an excellent base for a mid-gen Switch, if you ignore timing. The power consumption is right in line with the base Switch, it's got 2x the SMs as Mariko, and 6 CPU cores instead of 4. You could get Mariko level batter life with a little DLSS, or extended power across the board, for a cost about the same as the existing device.
Either Orin Nano was going to be late by 6 months due to capacity issues at the fab, pushing Switch Pro back, or the existing Orin Nano is a rapid replacement for the original device. Nintendo realized what we did, that a mid-gen refresh in 2023 wasn't a great move unless it could step into as a successor. The project was pulled, and what software wasn't cancelled was either left on Mariko, or moved to Drake, and Drake was moved to N4, which neatly explains every weird question we've had about the damn thing.
Regarding Mochizuki's reports, they changed over time, and
@mariodk18 posted a helpful summary
here. Up until March 2021, he simply stated that there would be a 4K-capable Switch released in 2021. In March 2021, he stated that it would have an OLED screen and use DLSS, and then in September 2021 he stated that a 4K-capable Switch wouldn't be released until late 2022 at the earliest.
My hypothesis (which could be entirely wrong) is that Mochizuki's sources were talking about a couple of different models, which were conflated together. An overclocked Mariko model may have been the basis for the first couple of reports, and that could have been cancelled (say in early 2021), and at the same time Nintendo started talking to developers about their DLSS-capable next-gen system instead. So in late 2020 he talks to developers about a 4K Switch launching in 2021 (the mid-gen upgrade), then in early 2021 he asks other developers and they say "The 4K Switch? Yeah, it's using DLSS" (but they're actually talking about the next-gen console that's further out), and then supply chain reports tell him Nintendo have ordered a load of OLED panels for delivery this year, and he connects the dots to get a 4K-DLSS-OLED system in late 2021.
I'll readily admit that an overclocked Mariko is very much a stretch to be called 4K-capable, but perhaps it was more "4K output" rather than rendering at (or even close to) 4K.
I didn't say that either! This is getting into the weeds a bit, I'm not trying to argue this definitely happened. I'm just looking at possibilities. Let me walk it through.
The original Nvidia roadmap for this year had Nano Next (not branded as Orin) launching in 2022 before Orin did. Before the Drake reveal in March, many of us speculated that Nano Next was Dane.
Late last year Nvidia moved Nano Next to after Orin on their roadmap. Orin Nano, the product we eventually got, is a floorswept Orin. In modern chip manufacturing, floorswept chips don't come before their full fat counterparts.
Why was there ever an expectation that Nano Next would launch before Orin? Did it, at one point, have a large customer with a timeline? Was it, at one point, an independent chip, rather than just a floorswept version of its Big Brother?
Just for clarification, the old roadmap had Nano Next to launch before
Jetson Orin, but not necessarily before Orin itself. Orin was shipping to automotive customers for quite a while before any of the Jetson products shipped (since sometime in 2021, at least). So conceivably they could have planned to take dies that didn't meet automotive bins and use them in Jetson Nano, although it would have been a bit strange to not have any intermediate binned chips (like they're using for Jetson Orin NX), unless some automotive customers would use these as well.
One thing that has me thinking Orin was always planned to be used for Nano Next is that it's just much better suited to Jetson use-cases. With double-performance tensor cores, the Orin Nano 8GB (with 8 SMs active) would comfortably outperform Drake's 12 SMs in ML workloads clock-for-clock. A hypothetical pre-Drake Nintendo chip would have almost certainly underperformed even the current Orin Nano 4GB, so wouldn't have made much sense to use for a Jetson product when significantly cut-down Orin chips are apparently viable.
The other, related issue is DLSS performance. It's hard to judge exactly how much performance is required for DLSS in a console environment, but I feel like they'd be stretching to get close to 4K with a chip the same architecture as Drake, but less powerful. Personally, my expectation for Drake is DLSS to 4K at 30fps, but only around 1440p at 60fps (which I'd be perfectly happy with). If there were a pre-Drake chip with, say 6 SMs, you'd then be looking at perhaps a 1440p max from DLSS at 30fps, and not exceeding 1080p at 60fps. The logical thing to do here would be to use Orin's architecture instead, with double-performance tensor cores, as a 6 SM chip could then get the same ML performance, clock-for-clock, as our 12 SM Drake. This is basically what I was predicting before the Nvidia hack, and I still think it would have made sense for a mid-gen upgrade, but if there was a change in architecture like this, I would expect some kind of remnants in the Nvidia hack pointing to it, such as a different code for the chip (ie not T239).
Also, if Nintendo were planning to release an Ampere-based Switch device in 2021 (a year after Ampere desktop launch), then after cancelling that, why would they stick with Ampere for a 2023 release? If you're already making architectural changes, and possibly node changes, why not switch to Ada? It would still be launching about a year after the desktop Ada launch, so it wouldn't be any more cutting-edge than Ampere in 2021.
I'm not 100% ruling out the intermediate chip theory, and I admit that the overclocked Mariko theory is far from perfect, but it just seems the less-improbable scenario to me at the moment.
