By using this site, you agree to our Privacy Policy and our Terms of Use. Close

Forums - Nintendo Discussion - What do you want in a Switch 2?

So this year maybe the Switch Pro (2TF) and next year probably the Switch 2 (4TF) or just skip the Pro and go for the Switch 2.



Around the Network
Cobretti2 said:
jonathanalis said:

From here: https://www.extremetech.com/extreme/301531-nvidias-new-jetson-xavier-nx-adds-horsepower-to-ai-at-the-edge
The jetson xavier NX is capable of 6 TFLOPS FP16 at 15W.
Scaling to FP32 and to 10W (linearly), it would be capable of 3 FP32 TFLOPS at 15W, and 2 TFLOPS at 10W.

Unless I am loosing something, and it do not scale linearly...

That is interesting. may need to do more checking up on this.

Wiki been saying 0.8TFLOPS
https://en.wikipedia.org/wiki/Tegra#Xavier

I assumed it was right as the X1 switch has is right.

That 6TFLOPS is for AI, I'd assume it comes from 48 Tensor cores, not CUDA cores. So indeed, at 10W (which assumes 800MHz clock), FP32 performance is around 0.6TFLOPS.

Similarly, 2080Ti has 107.5 TFLOPS from its 544 Tensor cores and 13.5 TFLOPS of CUDA FP32 performance.



HoloDust said:
Cobretti2 said:

That is interesting. may need to do more checking up on this.

Wiki been saying 0.8TFLOPS
https://en.wikipedia.org/wiki/Tegra#Xavier

I assumed it was right as the X1 switch has is right.

That 6TFLOPS is for AI, I'd assume it comes from 48 Tensor cores, not CUDA cores. So indeed, at 10W (which assumes 800MHz clock), FP32 performance is around 0.6TFLOPS.

Similarly, 2080Ti has 107.5 TFLOPS from its 544 Tensor cores and 13.5 TFLOPS of CUDA FP32 performance.

far out, talk about confusing lol.  two different measures 



 

 

The Tegra X2 hits 0.6 TFLOPS easily, lol.

Whatever Nvidia will have for 2022 for a Turing based Tegra will be way beyond that.

Just because they're not sharing a "roadmap" to people on the internet means nothing, Nintendo is the main customer for the Tegra line now and their needs/wants are what they will focus on and Nintendo wouldn't want that information out there because everyone and their grandma will know it's for Switch 2.

The Tegra X1 is 7 years old come March 2022, that's ancient, to say that they wouldn't have a chip that completely blows that away is laughable.

The Xavier is hard to measure because it's not designed for graphics compute applications, it's designed for self driving cars so the AI performance obviously is paramount. 



Soundwave said:

Powerful enough to run all PS4/XB1 tier games with ease, that should be a no brainer, but I would like to have performance that can accomodate even PS5/XB2 ports at 720p-900p.

The Switch really actually was not that far off from running PS4/XB1 titles as the Witcher 3 port illustrates (and the overclocking hacked Switch's show Nintendo had more power there that they could have tapped into). And there's far less overhead for the current Switch-PS4 ports because PS4 ports run at 1080p while on Switch they maybe can do as low as 540p undocked/720p docked. But for PS5/XB2 they're gonna have to run at 4K, which is quite a lot of their resources being eaten up by having to render that many more pixels.

I'd like Switch 2 to make ports like that easier, thus inviting more developers to try.

To that end I'd reccomend a new performance profile for undocked ... overclock mode that lets you run docked performance undocked if you have a battery pack.

Because really the reason the Switch has to cripple it's performance so much is because of the undocked mode running off a tiny battery. But extra battery's today are dirt cheap (one can get a 5000 MaH, doubling the Switch's battery life for like $15).

Even people who have modded Switches have tested overclocked performance undocked and the system doesn't get hot enough that its uncomfortable to hold undocked. It's just you have less battery life basically.

Well, duh. It will be much more powerful than a ps4, so of course it can receive any of it's games with ease.

You mean in portable mode, right? Because the switch 2 should have no problem getting multiplats with ps5 running in 1440p docked.



Around the Network
Cobretti2 said:
HoloDust said:

That 6TFLOPS is for AI, I'd assume it comes from 48 Tensor cores, not CUDA cores. So indeed, at 10W (which assumes 800MHz clock), FP32 performance is around 0.6TFLOPS.

Similarly, 2080Ti has 107.5 TFLOPS from its 544 Tensor cores and 13.5 TFLOPS of CUDA FP32 performance.

far out, talk about confusing lol.  two different measures 

Thats why its specific, that its only in A.I funktions.



Soundwave said:
Pemalite said:

What makes nVidia flops different or better than AMD's?

Xavier is built at 12nm, not 16nm.


The 12nm process is based on 14nm anyway, which in turn is based upon 20nm which it retains for it's BEOL.

Even so a 7nm Turing based chip is going to be a significant improvement from that. 

The FLOP Nvidia/AMD thing I think stems from AMD GPUs underperforming against Nvidia processors while claiming the same FLOPS or even more in some cases. They're saying this too with regards to the PS5 leaks, that that's 9TFLOPS or whatever relative to a Nvidia GPU like a 2080, but those flops are greater than 9TFLOPS of AMD GCN part (PS4 era). 

That's not telling us how AMD's flops are different or worst than nVidia's though, which is what I asked...

Cobretti2 said:
HoloDust said:

That 6TFLOPS is for AI, I'd assume it comes from 48 Tensor cores, not CUDA cores. So indeed, at 10W (which assumes 800MHz clock), FP32 performance is around 0.6TFLOPS.

Similarly, 2080Ti has 107.5 TFLOPS from its 544 Tensor cores and 13.5 TFLOPS of CUDA FP32 performance.

far out, talk about confusing lol.  two different measures 

Actually even more measures.
FP16, FP32, FP64, INT4, INT8, INT16, Rays/s, RTX Ops, and so on.

Then you have Texture and Pixel Fillrate and more...

Which is why anyone who uses just plain-jane teraflops are either not being comprehensive or just don't know what they are talking about.



--::{PC Gaming Master Race}::--

I'm assuming the Switch 2 comes out in 2024, and that it is mostly the same idea as the Switch.

The big thing I'm hoping for is an embrace of high refresh rates. I think that this will be one of the biggest trends for the next 2-3 generations. So I'm hoping that the Switch 2 is capable of 4K 120Hz when docked (I know that very few games will actually run at 4K 120fps, so I'm just saying that I want the Switch 2 to be able to output both. This way some games can max out frame rate and others resolution. What I'm trying to say is that I want Smash 6 to be 120fps.) and the handheld screen to be 1080p 90hz with an HDR OLED (or whatever else is as good and available by then), thin-bezel display. I really don't think that there's any substantial benefit of going past 1080p at that screen size, so it's much better to prioritize frame rate. By then, 90hz should be easy at 1080p. Maybe even 120? But with the concern of battery life I would be content with 90hz. Also I want the Switch 2 to output variable refresh rates both docked and in handheld mode.

I think that all these things are reasonable by then. The technology required to do all of these things exists today and should be quite cheap by the time the Switch 2 releases.



NNID: garretslarrity

Steam: garretslarrity

Pemalite said:

Cobretti2 said:

far out, talk about confusing lol.  two different measures 

Actually even more measures.
FP16, FP32, FP64, INT4, INT8, INT16, Rays/s, RTX Ops, and so on.

Then you have Texture and Pixel Fillrate and more...

Which is why anyone who uses just plain-jane teraflops are either not being comprehensive or just don't know what they are talking about.

Yer I get all that but when something comes out and the measure is say 2TFLOPS vs 9 TFLOPS, you can't expect miracles to happen with porting games. A lot of big developers will just say why bother, too much effort. I know there is other measures and factors, but such a big difference can't be overcome with other factors.

That is why I keep saying I am happy with the Switch as is, no point seeing a Switch 2 in 2022 if all we will get is the same output of games (mainly Nintendo and indies), if that gap is too big. May as well wait till 2024 and build something with Nvidia specific for Nintendo.



 

 

Cobretti2 said:
Pemalite said:

Actually even more measures.
FP16, FP32, FP64, INT4, INT8, INT16, Rays/s, RTX Ops, and so on.

Then you have Texture and Pixel Fillrate and more...

Which is why anyone who uses just plain-jane teraflops are either not being comprehensive or just don't know what they are talking about.

Yer I get all that but when something comes out and the measure is say 2TFLOPS vs 9 TFLOPS, you can't expect miracles to happen with porting games. A lot of big developers will just say why bother, too much effort. I know there is other measures and factors, but such a big difference can't be overcome with other factors.

That is why I keep saying I am happy with the Switch as is, no point seeing a Switch 2 in 2022 if all we will get is the same output of games (mainly Nintendo and indies), if that gap is too big. May as well wait till 2024 and build something with Nvidia specific for Nintendo.

Porting isn't just about flops...



--::{PC Gaming Master Race}::--