How Will be Switch 2 Performance Wise?

[Soundwave]

sc94597 said:

Pemalite said:

1. It is a false statement, it's not a "Hand me down". - Nintendo made a purchase that was disconnected from the Shield lineup.

As for the CPU efficiency cores being disabled, that was NOT a Nintendo thing, that was an nVidia thing.
Because devices like the Pixel C also had the ARM A53 cores disabled as well.

nVidia eventually removed those cores entirely in later revisions of the Tegra X1 SoC's that ended up in Shield devices.

As for T239, it's a cut down T234.
It's literally nothing special, it's literally not a new or unique design.

It's a die-harvested part.

sc94597 said:

2. U.S companies are the leader in chip design (although not exclusively) and the design work does happen in the U.S. Why do you keep bringing up manufacturing and fabrication? Nvidia is not a chip manufacturer, they're a chip design company. Nintendo's contract with Nvidia is for chip design, and you know this because you say as much in your response here. The T239 would not exist if Nintendo didn't work with Nvidia to ask for it to be designed for them. Tegra Orin would just be a SOC family with one fewer SOC. The T210 would exist if Nintendo didn't contract with Nvidia for the Switch. That's the difference. That is what is meant by "hand me down." 

2. Fabrication is the basic fundamental cornerstone that determines what the design of a chip can be.

Certain fabrication processes have strict die-sizes, voltage limits and frequency curves that need to be strictly adhered to otherwise issues such as electromigration come into play.

You cannot take a Geforce RTX 5090 with it's 92~ Billion transistors and drop it onto a 14nm process node and call it a day, it doesn't work that seamlessly, sacrifices and changes to things like placement of dark silicon, pipelining and more need to be done to reduce things like crosstalk and leakage.

3. The T239 could exist without Nintendo, it's a Die-harvested part.

You have the T234 as your full blown chip. - nVidia then reduces it down by disabling parts of the chip to meet yields.
In-fact they did it with Orin 32GB which lost 256~ CUDA cores.

And then they did it again with Orin 16GB which was 50% of the CUDA cores.

And then they did it again with Orin 4GB which was 25% of the CUDA cores.

These are all the same chip. The same design.

Nintendo isn't getting a special or unique (Other than being a variant of already designed ORIN chips.) chip here.

There is no secret sauce.

4. You can take the T234 and edit the firmware to disable CUDA cores to match the T239 chip. - Making it identical.

Tegra is used in a lot of different markets, you just need someone who wants to use a cut down chip and nVidia would oblige, it's not exclusive to Nintendo, it's exclusive to who has a wallet and will pay enough, nVidia holds the rights to the chip and it's technology, not Nintendo.

There is no specific design choices that make it unique other than it being a die-harvested part.

sc94597 said:

5. And Tegra T234 is not the chipset in the Switch 2. The Tegra T239 that is in the Switch 2, is a different SOC from the T234. It's not the same situation with the original Switch where the same exact chipset (T210) was popped into the Switch 2, the efficiency core cluster deactivated (not even removed) and frequencies changed. The T239 has specialized accelerators for gaming-specific tasks, different SM/core counts, and has the specialized edge-compute accelerators (NVDLA) removed compared to the T234. You don't see anything like this with the T210 vs. T210 (in another device) situation because that was an actual general-purposed chip meant to be used in tablets (amongst other devices) and not specifically designed for edge-compute as the T234. Nintendo was able to just purchase the T210 from Nvidia as-is, because it worked well enough. They couldn't do that with a T234 (for its size alone.) And even then Nintendo had to disable its efficiency core cluster that existed physically in the Switch, which they wouldn't have done if they had a chipset designed specifically for the Switch, as they do with the T239 for Switch 2. They'd just not include them from the start and have a different CPU configuration that is fully utilized by the Switch, if the Switch's chipset were semi-custom like the Switch 2's. This is why the T239 doesn't physically have the NVDLA; it's a semi-custom chip for the Switch 2, whereas the T210 was not for the Switch and it therefore had redundant/vestigial hardware (like the A53 cluster.) 

5. Correct. T234 is not. But it is based on T239.

What specialized accelerators for gaming specific tasks? Got evidence?

NVDLA or "nVidia Deep Learning Accelerator"  is open source and can run on Switch 2 or any Tegra/nVidia powered device, it's up to developers to include it.
https://en.wikipedia.org/wiki/NVDLA

1. It is "hand me down" in the way I was using the idiom and for the context of the discussion I was having with Bonzobanana. You interpreted that phrase to mean that I was suggesting Nintendo got a bunch of physically existing chips left-over from the Nvidia Shield. I clarified multiple times that the design is what is being "handed ... down" not the physical units and that is precisely what I was intending to mean when I used the idiom. 

2. Okay, who said that fabrication doesn't matter when it comes to design? That was never a point I made. The point being made is that the SOC designers work for Nvidia and are making high salaries on the order of six and seven digits. Any extra work they do designing new bespoke SOCs for a special customer, who is the single purchaser of that SOC, has additional costs. They're nowhere near the costs of designing a new micro-architecture or a SOC family, but these costs are still significant. 

3. It could exist (in some form), but it wouldn't. The point being made isn't that it could exist. The point being made is that it wouldn't exist. That is evidenced by the fact that Nintendo is its only purchaser and is almost certainly involved in the design process of the chipset as that sole purchaser by setting the constraints of what they want the chipset to be able to do. Again, the specific SOCs STILL HAVE DESIGNERS, ENGINEERS, and TESTERS who have to make them fit their target device. These decisions aren't magically made without employees involved. Iterative testing still needs to be done. Project managers and planners still need to be involved. Teams need to come together to do that planning, implementation, and testing. All of that is internal to Nvidia. All of that costs something. 

4. Okay, and how are you going to fit that 455mm² chipset in a Switch 2 sized handheld? What do you cut with meeting your single consumer's (Nintendo's) performance and efficiency goals? Who decides this? Nvidia employees, maybe? Without the different CPU of the T239, is it a T239? Remember the T239 isn't just its GPU, but the whole chipset, including its CPU as well. 

Nvidia could indeed sell the T239 to other customers, but nobody else wants it (as otherwise they would have bought it), and it doesn't change the fact that they are designing the chipset, initially, for a single customer of that chipset with an intended use-case in mind. 

5.

T234 has the NVDLAs built into its SOC (it has two of them vs. one in the cut-down variants) while the T239 doesn't have one at all. This had to be decided by real employees at NVIDIA when designing the T239, not Samsung, not solely Nintendo. 

It's also known that the T239 uses different CPU cores (either A78C or a hierarchy of X1/A78C/A55, depending on which rumor is true vs. exclusively A78AE.) So no you couldn't just cut or under-clock the CPU and have a T239. Orin NX, Orin Nano, etc retain the A78AE cores. 

The addition to the T239 that didn't exist in the T234 (and its cutdown variants) we know of is the file decompression block. 

It isn't even certain that the T234 and T239 were designed by the same team. 

https://wccftech.com/nintendo-switch-2-8nm-soc-ampere-efficiency-goals/#:~:text=Additionally%2C%20the%20T239%20chip%20may,file%20decompression%20block%20required%20for

"Additionally, the T239 chip may have begun development as some sort of offshoot of the much bigger and more power-hungry T234 chip, but the two chips have been designed by different teams, have different CPUs, and the Nintendo Switch 2 chip has some additional features, such as a file decompression block required for fast loading and asset management, so it's not a cut-down version of the chip used in the automotive segment."

There is no game console that doesn't have a "hand me down" chip architecture to some degree. 

AMD and Nvidia will never make a chip expressly only for a freaking peanut margin, low volume (yes, even 150 million units sold in 8 years is nothing special in the broad sense of electronics) game console. 

The PS5 and XBox Series S/X are hand me down chips too. 

That more relevant point for the Switch 2 is how efficient the chip is. Any idiot can put a mobile 4080 GPU into any "mobile" system, there's nothing special about that, it's about what kind of performance can you pull given a tablet sized form factor that has to be portable. 

A Tegra T234 is designed to run off a massive freaking car battery. 

What I think is going to be the fundamental difference between Switch 1 (Tegra X1) and Switch 2 (Tegra T239) is the T239 can handle PS5/XBSS ports more easily than the Tegra X1 could handle PS4/XB1 "impossible ports". We see that already there is a growing list of PS5-XBSS games already announced or stated as coming from reliable leakers. Star Wars Outlaws, AC Shadows, Wild Hearts, Microsoft Flight Simulator, Final Fantasy 7 Rebirth, Cyberpunk 2077 w/Liberty City DLC, etc. etc. Nintendo's hardware team has stated they weren't satisified with the Tegra X1 and wished it was more powerful as it couldn't handle all types of games ... it would seem they feel more confident this time that the Tegra T239 can handle more of just about anything thrown at it.