I don’t know. We can theorize lots of things. Absence of evidence is not evidence of absence.
This is a thing I say a lot. My grandfather would reply with things like, "how do you prove a negative." You can't - unless you can prove a counterfactual. I can't "prove" a counterfactual, but I can provide my existing evidence. I think that evidence tells a compelling story which sets the burden of evidence for a competing theory. Presented as a timeline
March 2018
Nvidia Orin, the next generation Tegra SOC targeting automotive systems is announced
July 2019
Switch Lite announced. It runs a version of the Switch's TX1 chip updated for Nintedo's needs - it is die shrunk for better power, the Tegra Security Module, a standard part of Tegra SOCs, is removed as it is ironically the basis of Switch exploits.
August 2019
A verification engineer at Nvidia states that they are working on Orin and that they we working on a "File Decompression Engine for games".
July 2020
Nvidia posts a job listing for a Tegra engineer for game consoles to work on optimizing DLSS 2 for that particular space
June 2021
Nvidia leaker kopite7kimi claims that the next Nintendo console will feature a modified version of Orin, called T239.
Internal commits on Nvidia's linux repository begin to contain T239 drivers.
These include references to the File Decompression Engine, which is not a part of base Orin.
The net picture at this point is that since 2019 Nvidia has been working on a chip for a game console based on Orin, and that T239 is that chip. Are there others? Maybe. Let's continue
September 2021:
Bloomberg/Mochizuki drops the famed "11 developers" article, a direct response to their admittedly bungled reporting on the OLED and Nintendo's aggressive retraction. It specifically mentions DLSS and that devkits are in the wild.
There are three possibilities. One is that the new device is based on Tegra Xavier, as the only SOC on the market at the time that could do DLSS. The second is that it is based on Orin, then upcoming which is also capable of DLSS. The third possibility is that the new device is a totally new concoction, unrelated to any of the existing SOC designs from Nvidia.
March 2022
NVN2 source code is leaked. It explicitly states it runs on T239, has DLSS and RT, and explicitly states several parameters of the underlying hardware, which are in line with a cut down version of Orin's GPU. There are changes indicating updates to comments and internal documentation updating references to Maxwell (the Switch's architecture) to Ampere (Orin's architecture). Critically, there is no indication that these changes passed through an intermediary "Volta" variant (Xavier's architecture)
Also of note - the DLSS integration is clearly new as there are comments in the code referring to not being sure how to change from a PC programming model to the one required by Nintendo.
April 2022:
Commits to Nvidia's internal Tegra Linux repo contain references to fixing PCIe timing bugs on T239, implying that they're running on real hardware.
May 2022:
Orin's release confirms nearly every leaked detail about it's design, many from kopite7kimi who is the original source for the (confirmed) T239 leak
August 2022:
Nvidia quietly drops new public code and documentation as part of their open-gpu-docs project on Github. It is the first public mention of T239, documenting it as an Ampere GPU right beside T234 (the base Orin). Notably, T239 is a listed but unannounced product. No other unannounced products or SOCs are mentioned.
September 2022:
Nvidia pushes some internal Linux code upstream to Linus's branch. This includes a T239 CPU driver, which specifies that T239 uses a single 8 core CPU cluster. There is only one viable CPU that can do that, the A78C. Orin uses the A78AE, a variant of the A78 intended for automotive work. A78 would be the general purpose version. A78C is a high performance version designed for "mobile gaming".
Additionally, external partners who mirror the Tegra 4 Linux project receive updates around this time from Nvidia's internal Linux development that contain multiple T239 references, which paints an extensive picture of the hardware, including the FDE
Notable - commit history indicates that Linux development began prior to the NVN2 hack, there is no indication of a major rethink, and none of the details conflict with the NVN2 hack's hardware details.
The facts so far:
Nvidia is developing a chip called T239 for Nintendo, it is based on Orin, and has a well known design confirmed by both illegal hacking, public documentation, and leaks, all of which match each other. That chip matches the specs expected of devkits that were reported on in 2021 (DLSS/4k capable), and is ready to be manufactured, if it isn't already being made.
The Lite reflects a separate custom chip design for Nintendo. Orin design work started before the Lite was launched, likely well before. A game console based variant of Orin began design in 2019 at the latest. There was no
PascalVolta/Xavier based chip for Nintendo that reached a point of NVN development.
The counterfactual burden:
If T239 isn't the next Nintendo chip either it's slated for a later device, or it's not coming to Nintendo devices at all.
It is hard to conceive of what device t239 would be targeting if not the "next" device, as it is both clearly designed for Nintendo and ready to come off the line, and likely conceived for a process node that doesn't have many years left. Nvidia would be taking a Nintendo designed chip, with multiple pieces of custom console destined hardware (the FDE, the CPUs) and putting it in other products for a few years while waiting for Nintendo to catch up?
And considering T239 is getting public documentation dumps and Linux updates, the product hasn't been scrapped. If it's not coming to a Nintendo device at all, then 3 years of design effort went into the chip, which meets no existing market need for Nvidia that isn't met by Orin, and then Nintendo pulled out in the last year, Nvidia decided to make the chip anyway, chasing the sunk costs of the design with the actual costs of manufacturing. Nvidia doesn't charge Nintendo anything for this contractual break, and/or Nintendo hides the huge sunk cost and contract breakage fees as "R&D."
Are these events possible? Or perhaps a third scenario I can't think of? I suppose so, but you can maybe see why I'd go 10:1? I can read the source code, that might make this story extra compelling to me, I can see how far along the driver work is in the Linux kernel, and in the various bits and bobs shared of the NVN2 code. I'm not doubtful about the state of this code as some others processing the info second hand might be.
Okay but:
What if the design of T239 doesn't match the state it was in when the Linux code dropped and the NVN2 hack happened?
The Linux code, I don't buy it. Nvidia is pushing these updates to partners and into the mainline kernel, and in places where T239 shares drivers with existing hardware. The updates are pretty minor, but pushing through the review process to get code into upstream that might break production hardware in order to support an unstable, unreleased product is something developers get Linus yelling at you for. It is also a superb waste of energy.
That leaves some of the GPU specifics. It's definitely Ampere, that's documented elsewhere. The only real place that I think is possibly reconsiderable is the number of SMs. Yes, yes, I am aware that it is hardcoded into the NVN2 source code, but it would have to be based on NVN2's design. That doesn't mean that they couldn't change it if the hardware failed to deliver.
But they'd know the performance/power ratios at the time of the NVN2 hack, because Orin was manufacturing. Orin has 14 SMs - a "first draft" design would probably start at 14, or move much lower if power/perf ratios weren't right. 12 implies they were already dialing in the exact size of the GPU.
As does the fact that this isn't a test ROM or a benchmark tool pushes that further. This isn't a Windows implementation for an early SDK, this is a port of the entire graphics API to new hardware.
It seems like a change from 12 to, say, 8, wouldn't be about failing to meet certain power/perf metrics, at this point, but about changing power/perf goals. I'm dubious, but I will admit there is more room here than anywhere else.
” as it doesn’t tell me there is a reason. It just tells me it was done for shits and giggles.
