He got the motherboard on the Chinese equivalent to eBay (Xianyu) for $150

Resume:

207mm² v.s. 118mm²

It's also larger than the RTX 3050 Ti (mobile) GPU, Apple M2, Ryzen 7840H (Z1 extreme) and Qualcomm's X elite.

Tape out year of the SoC was on 2021 as annotated on the die. They also looked at the RTX 3050 GPU die, which had a tape out year of 2020.

T239 brief summary.

8 Core A78C cores, with 4MB L3 Cache and 256kb L2 cache on each die

The die area of the A78C is the same as the 8nm T234 orin, indicating the same process node.

The C and AE suffix doesn't bear much meaning on the core design itself, but rather the ARM classification indicating Compute and Automotive and how the cores operating

6 TPC, 2SM/TPC, 1536 CUDA Cores

Ampere design, although it's layout seems odd as the TPC's are separated

The design of the GPU seems more alike to Ada Lovelace

in terms of area of the TPC, the T239 is 22% smaller than the orin T234

There are some differences in the chip design itself, it's not exactly 8nm and some aspects are closer to Samsung's 10nm process node fabrication.

He highlight the previously "leaked" clocked speeds and like DF mention that its closest equivalent is the RTX 2050, although heavily downclocked.

Like DF they also do a simulation of Switch 2 performance with the RTX 2050 downclocked as a proxy.

The docked performance is aligned to the GTX 1050 Ti in docked mode and GTX 750 ti in handheld mode in Synthetic tests Steel Nomad Light

It's far from the Xbox Series S and the closest mobile equivalent are Apple's A18 Pro, M1 in docked mode, and the Steam Deck is a bit better in handheld mode * They test Cyberpunk 2077, Black Myth Wukong, Monster hunter Wild, KCD2

They also took a T234 board (Orin) and tested CPU performance through synthetics (geekbench 6)

Downclocking it to 1.1 GHz and 1 GHz, its single core performance is twice as much as the PS4 and 3x the Switch, closest laptop equivalent is the i7-4700HQ.

Closest mobile equivalent in multithreaded performance is the A12 in the iPhone, which btw has more than 2x the single threaded performance compared to the "Switch 2"

They also used a downclocked i7-10700KF to approximate Switch 2's CPU performance

Some Benchmark Results:

Cyberpunk 2077

Docked: Quality Mode (1080p, Dlss Quality, Low Preset) 30-35fps, Performance Mode (1080p, Dlss Performance, Low Preset) 35-40fps

Handheld: Quality Mode (1080p, Dlss Performance, Low Preset) 28-30fps, Performance Mode (720p, Dlss Performance, Low Preset) 35-45fps

Black Myth Wukong (First battle, that is pretty heavy)

Docked: (1080p, Low Preset, Dlss Balanced) 30-35fps, (1080p, Low Preset, DLSS Ultra Performance) 40-45fps

Handheld: (720p, Low Preset, Dlss Ultra Performance) 20-30fps

Monster Hunter Wilds

Docked: (720p, Low Preset, Dlss Ultra Performance) 20-23fps

Handheld: (720p, Low Preset, Dlss Ultra Performance) 15-18fps

Kingdom Come Deliverance 2

Docked: (1080p, Medium Preset, Dlss Performance) 30-40fps

Handheld: (720p, Medium Preset, Dlss Performance) 30fps

Call of Duty Black Ops 6 & Warzone

Docked: (1080p, Low Preset, Dlss Ultra Performance) 50fps+

Handheld: (720p, Low Preset, Dlss Ultra Performance) 40fps+

Full Video:

Samsungs 8nm is literally based on their 10nm process. - The claim the Switch 2's SoC is "close to 10nm" is thus a laughable statement, because Samsung 8nm is not 8nm geometrically, it's a refined 10nm process, advertising at it's finest.
https://fuse.wikichip.org/news/1443/vlsi-2018-samsungs-8nm-8lpp-a-10nm-extension/

Nothing here is really new information, we already knew it was a die-harvested Orin part.

The Synthetic benchmarks are not representative of real-world performance.

Samsung's 8LPP/LPU process has a tighter gate pitch, different contacts, and different track and cell heights. It's an extension of 10LPE/LPU but it's not like they slapped a new name into the same old thing.

Granted, the custom 8N process Nvidia used for consumer Ampere already lacked some of the features of Samsung's "8nm" node, but the Switch 2's SoC (which is not a die-harvested Orin, by the way...) even more so.

That is essentially how all these "updated" nodes work. Even in the same "Process family". I.E. 7nm vs 7nm+ at TSMC has a 1.2x improvement in density for the  transistors.
There are normally changes to pitches, contacts, cell heights and more to fit various targets like frequency, leakage and so forth.
https://community.cadence.com/cadence_blogs_8/b/breakfast-bytes/posts/tsmc-oip-17

The Switch 2 is likely the exact same chip as the Tegra Orin NX 16GB - Because there needs to be a few redundant CUDA cores for yield purposes as the Switch 2 SoC is over 200mm2.
128 CUDA cores is an SM, so having 2x SM's in "reserve" wouldn't be a stretch.

They compare  the portable Switch 2 to a Geforce GTX 750 ti which is about 1.4 Teraflops but they are talking about 17-19W power yet the Switch 2 only has a 20Wh battery and with a battery runtime at a minimum of 2 hours it can only use 10W per hour of which half that will be taken by the screen approx. So we are talking 4-6W per hour. They haven't factored in the small battery capacity into their analysis.

Also when Intel decided to call their 10Nm process 'Intel 7' it was due to Samsung's description of their process being 8Nm and yet Intel's 10Nm had greater transistor density and power efficiency so they came up with 'Intel 7' just to state their process was superior. Basically the Samsung 8Nm process is mainly 10Nm. I'm not sure how the TSMC process compares, I'm guessing not as over-stated as Samsung but not sure.

I've still yet to see any explanation on how the Switch 2 can provide 1.4 Teraflops of GPU performance in portable mode with such a fabrication process and only 4-6W per hour? Surely the 1.4 Teraflops figure is theoretical maximum power for short bursts nothing more. The reality will be much, much lower as an average. Trying to compare to original Switch it was claimed for a while that portable was 157 Gflops but some claims have it as just over 200 Gflops but from what I've read the reality was 90-140 Gflops in actual use based on what hackers/modders etc claimed. So in theory if the Switch 2 follows the same path it will be a lot lower in performance in reality than its theoretical maximum performance for portable mode (not docked).

Maybe this is why Nintendo are holding back sending journalists retail Switch 2's until launch day as journalists will be a little disappointed in overall performance. Yes I totally accept they are probably holding back units due to poor Switch 1 compatibility on launch hardware but it could also be performance factors too.

Poor dated/cheap fabrication process that is power hungry with small capacity battery. This is still what I'm waiting to be addressed in the Switch 2 spec. I've looked at loads of articles and yet nothing seems to be stated with regard how this works out. I still believe the Switch 2 Gflops figure for portable use will be far below 1 Teraflop in order to provide that 2 hours minimum runtime. I'm certainly happy to be proven wrong though.

Third party is not worth it, as expected.

Geekerwan did an excellent job, he is a contender for DF and they seemed happy with his vid aswell