Yeah, when I use the Jetson Power Estimator tool with a 14 SM Orin chip, 8 CPU cores clocked @1113 Mhz max (according to leak Switch 2's CPU is about 1.1 Ghz), DLA's turned off entirely, all non-SOC board features turned off. CPU and GPU load set to "low" and GPU max clock at 816Mhz (to accomodate for the fact that Switch 2 only has 12 SM, 14SM @816mhz ~ 95% the performance of 12SM @1007.3Mhz.) 

When you do that, the estimated power draw is 18W, which is about 20-30% higher than Switch 1. 

So these are the explanations: 

1. The Switch 2 is running at a somewhat higher TDP docked than the Switch 1 (18W-22W vs. 11-15W.) Nintendo had some room in terms of thermals in Switch 1, evidenced by home-brew Switch overclocking, so that could be possible. 

2. There are some efficiency gains with T239 vs. T234, which could be as simple as the A78C cores being more power efficient at lower clocks than the AE cores, or could be Samsung 5nm. 

3. Some combinations of 1 and 2. 

or 

4. The leaked clock-rates are wrong. 

If we consider handheld mode is suppose to max at 560Mhz according to the leak, a comparable setup (performance-wise) would be 480Mhz @14SM, which the Orin power tool implies would pull about 13W-15W. That seems way too high. The Switch 2's battery has 16Wh capacity, that would imply a battery life of about 1.2 hours if fully utilized.

We know the Switch 2 has a low-end battery life of about 2 hours, so handheld mode should be about 8W overall, and more like 5-6W if you consider the screen, controllers, storage, etc are using power. 

Heck even if you choose the Orin NX (which only has 8SM) and clock the GPU cores to 408Mhz one gets 8W of power consumption for about 50% of the theoretical performance of Switch 2 handheld mode (12SM @ 560Mhz.) 

Now of course Nvidia might be giving a very liberal upper-range power consumption estimate to be safe, but still the difference at the very low frequencies seems quite large. Too large for there not to be some missing efficiency advantage for T239 vs. T234. 

A lot of assumptions and speculation here, which goes to show we don't know much about the specifics, even though it is clear the ballpark of where Switch 2 lands. 

Do you guys think we'll be able to record more than 30 seconds of gameplay at more than 720p30fps with the capture button? Switch 1 was very limited in this regard, so I hope the Switch 2 will feature a better capture functionality.

Dare I say it but Nintendo are pretty cheap and re-designing the ampere architecture for 5Nm surely would be quite expensive and time consuming. It was designed for TSMC 7Nm or Samsung 8Nm. Surely the whole point of DLSS upscaling is they can go cheap with the fabrication process as the console relies on upscaling to perform well. The fact that cyberpunk is rendering as low as 360p before upscaling means 8Nm is a perfectly satisfactory fabrication process. However I'll admit 5W per hour is quite low for portable mode. I'm so used to new Nintendo console launches having lots of forum posts about the consoles being a lot more powerful than they actually are. I remember the wii u had all sorts of claims for its GPU performance but in the end it was about 176 Gflops if I remember rightly and had to be based on its fabrication process and power consumption. However for a long time people were claiming figures like 700 Gflops and then 352 Gflops etc. Often with Nintendo the lowest spec wins the day. The Switch 2 can still manage incredible games thanks to its DLSS upscaling.

Maybe the mid-gen Switch 2 update will have a 5Nm fabrication process and a OLED screen etc and maybe boost battery life by 50-100% but normally Nintendo goes pretty cheap on launch hardware as they try to recoup development costs as quickly as possible. I'll be surprised (and impressed) if Nintendo has coughed up for a 5Nm re-design for the launch model. I think Nintendo always want the option of lowering pricing if the launch doesn't go as well as hoped. I suspect the margin on the hardware is excellent except perhaps in the US market now. The price in Japan is around $330 I think or less than 300 pounds if you are a UK citizen like me. The Japanese price would indicate a cheaper spec even if launching elsewhere at a much higher price because they think that is what the market will bear. 

People need to stop with the "normally Nintendo would ..." shit. 

This isn't Nintendo of 20 or even 10 years ago. New people are running the hardware division. The marketing is pushing 4K and 120 FPS of all things, the system is called Switch 2. It's not cheap. Nothing the "normally Nintendo would" crowd has said has basically come true. Almost everything about this system is different than that crowd expected. The Japanese yen is super low, that's why Nintendo can get away with a low price there, it's more expensive for every other market. 

Also Final Fantasy VII Remake SERIES confirmed for Switch 2, that will be Rebirth from the PS5 and the third installment most likely too from PS5. LOL, put that in your pipe and SMOKE it for the people who said even FF7 Remake Integrade wouldn't be possible on a Switch 2, LMAO. 

And also it looks way better than the PS4 and PS4 Pro versions, get that shit outta here. 

There's an excellent teardown on YouTube of some Chinese dude who got a Switch 2 motherboard. It looks like a mix of Samsung's 10 nm and the tighter pitch of the improved 8 nm node, surprisingly enough, maybe to optimize its power-performance curve at the lowest end.

The die design is a bit strange since a couple of SMs on the GPU are separated from the others. The design tape-out was in 2021, indicating it might be true that it was originally planned for an earlier release.

Like DF, they also simulate 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 the GTX 750 Ti in handheld mode in the Synthetic tests Steel Nomad Light. It's far from the Xbox Series S, and the closest mobile equivalents, graphics-wise, are Apple's A18 Pro and the M1 in docked mode, and the Steam Deck is a bit better in handheld mode.

Basically consistent with what we expected. Half-tier above base PS4 and half-tier below Series S. 

Series S performance is 1.7 times simulated Switch 2's. 

Simulated Switch 2's performance is 1.8 times PS4's. 

Series S performance is 3.11 times PS4's. 

Handheld mode is roughly Steam Deck level, but DLSS will be a big boon for image quality. 

The caveats being:

1. this is simulating Switch 2. CPU, memory hierarchy, etc can make a Switch 2 and RTX 2050 with the same theoretical performance, perform differently. Isn't clear which would have an advantage where. 

2. Synthetic benchmarks perform very differently from real games. Still, Steel Nomad Light seems to roughly correspond with the average modern game on low settings. 

Edit: Caveats also apply to the Series S and PS4 equivalents. A RX 6600 isn't quite a Series S in terms of memory hierarchy, and likewise for an HD7870 (drivers start becoming an issue for this old GPU.) 

Well, nice to see my napkin math turns out right where it supposed to be - I had it at 94-95% of 1050Ti, this chart has it at 96%.

They used downclocked 7870 for PS4, I used R7 265 (7850 with slightly higher clock), which has almost identical specs when it comes to Pixel/Texture fillrates, FLOPS and memory bandwidth, so my estimate from real world game benchmarks came about 1.6x PS4. Truth is probably in the middle, 1.7x, given that downclocking 7870, IIRC, makes geometry setup engine slower than PS4's, and going with R7 265 makes geometry engine bit faster than it should be.

Yeah, so it seems that using the Jetson Power Estimator Tool only made sense with the assumption that this is just a cut Orin chip with a different CPU (A78C rather than A78AE.) 

Looks like it is indeed not just a "cut down Orin T234." Each SM is smaller than a T234 SM.  

So it's not just the configuration that is custom here. 

Looking at this chart, it is clear that we have three tiers of performance that developers can target this generation. With Tier 3 representing low PC game settings, Tier 2 medium, and Tier 1 representing high/ultra. In terms of primary platform holder sales numbers Tier 3 ~ Tier 2 > Tier 1.

Tier 1: Higher-end PC hardware

Tier 2: Mid-end PC hardware/PS5/Series X

Tier 3: Performant iGPUs/Series S/Switch 2 Docked/maybe Switch 2 handheld for a while

(PS5 Pro is either at the upper end of Tier 2, similar to Series S in tier 3, or very low end of Tier 1.) 

Switch 2 handheld might squeak by into the lowest end of Tier 3 with M1 Apple Silicon and current generation mobile APU's. If Switch 2 misses out on any games due to performance reasons, it probably will be games released later in the generation and because of handheld mode falling behind. By then, Steam Deck players probably will be on a Series S-ish Steam Deck 2. The other PC handhelds aren't numerous enough to really matter. 

Who knows though, maybe newer DLSS revisions will mitigate this. I suspect the 10th Generation (if there even really is one) will go in hard on neural-rendering and we might see rasterization slowly become less and less important over this generation with gap-technologies being created to still enable reasonably performant cross-generation titles. 

Professional Ampere has much more cache and additional logic related to the tensor cores and FPUs than the consumer parts, so it makes sense all of that stuff would be removed to cut down on the die size. 

I wouldn't be surprised if the actual SM would have been even smaller than the traditional consumer parts; console manufacturers do that sort of stuff all the time.