Discussion Nintendo Switch and potential alternative ODM vendors

[Ghostsonplanets]

There are many colorful analogies involving bulls, accordions, or submarines among other things that I could make, but the SoC in the Switch successor will not be that powerful. It will have fewer CUDA cores and a lower clock speed than a 3050, which we all know is a ray tracing powerhouse.

The problem is that you're seeing by the optics of PC RT titles. RT is very scalable and very broad. It also has applications within the raster pipeline.

There's tons of R&D being poured by mobile vendors, iGPU vendor (Intel), etc for more scalable and efficient RT

Of course we won't be seeing Path Tracing or RT extensively being used at the level of Desktop. But the RT Cores will definitely have their applications on Switch 2, specially by virtue of being inside a console environment.

And Switch 2 will have even advantages over Mobile SoCs due to a much bigger GPU (1536 FP32), more bandwidth (120 GB/s for Switch 2 x 68 GB/s for premium mobile), more RAM available (11GB for applications), etc. So it's decently equipped to take advantage of Mobile RT solutions R&D.

Had Nintendo thought of RT units as unusable, they would have fused them off nor would Nvidia have written API calls for RT within NVN2.

 

[Mopetar]

Is the Switch 2 even using custom silicon? Last time they just used a Tegra chip that anyone could buy. If that's the case again the hardware would be there regardless of whether Nintendo wanted to use it or not.

They could certainly afford something with some customization, particularly after the success of the Switch, but they don't need to. Nintendo hasn't cared about having the flashiest graphics for several generations now.

I don't understand why you're comparing this to mobile phones. That doesn't matter. This is a chip that will have fewer cores than a 3050 and run at a lower clock speed. It's also using older generation ray tracing hardware that'll be two generations behind mainline Nvidia tech by the time it comes out.

If Nintendo can get some use out of the hardware that's great, but I wouldn't expect much. Hardware being used in a console doesn't magically make it capable of extraordinary performance. For reference the 3050 has 2560 cores and a base clock of 1.5 GHz. You're looking at 60% of the cores (assuming you're correct) at under half the clock speed.

 

[Saylick]

Is the Switch 2 even using custom silicon? Last time they just used a Tegra chip that anyone could buy. If that's the case again the hardware would be there regardless of whether Nintendo wanted to use it or not.

They could certainly afford something with some customization, particularly after the success of the Switch, but they don't need to. Nintendo hasn't cared about having the flashiest graphics for several generations now.

I don't understand why you're comparing this to mobile phones. That doesn't matter. This is a chip that will have fewer cores than a 3050 and run at a lower clock speed. It's also using older generation ray tracing hardware that'll be two generations behind mainline Nvidia tech by the time it comes out.

If Nintendo can get some use out of the hardware that's great, but I wouldn't expect much. Hardware being used in a console doesn't magically make it capable of extraordinary performance. For reference the 3050 has 2560 cores and a base clock of 1.5 GHz. You're looking at 60% of the cores (assuming you're correct) at under half the clock speed.

More semi-custom than fully custom, imo. It’s based on the T234 SOC but with a bunch of unnecessary stuff stripped out and some added customizations for gaming purposes. Eurogamer had a good article on it: https://www.eurogamer.net/digitalfo...dware-for-nintendo-what-is-the-t239-processor

 

[Ghostsonplanets]

Is the Switch 2 even using custom silicon?

Yes. T239 is based on T234 (Orin) blueprints but with extensive modifications to fit a gaming device.

If that's the case again the hardware would be there regardless of whether Nintendo wanted to use it or not.

It wouldn't. It's done because Nintendo wanted it. Wouldn't exist otherwise.

I don't understand why you're comparing this to mobile phones.

Switch benefitted a lot from Mobile R&D at the time, specially on FP16. Switch 2 will be a low power device and thus R&D done for accelerating performance on power and performance limited devices can be also applied for Switch 2.

It's also using older generation ray tracing hardware that'll be two generations behind mainline Nvidia tech by the time it comes out.

This older generation RT hardware is still extremely capable. AMD current uArch can only match it on Desktop equivalent GPUs. It's years beyond anything currently existing on mobile.

Hardware being used in a console doesn't magically make it capable of extraordinary performance.

It does give developers access to low level APIs and tools that can instead be turned into fine grained optimizations that can't be replicated on PC. Pre-building BVH or pre-baking shaders, better exposing the hardware, etc.

 

[soresu]

It does give developers access to low level APIs and tools that can instead be turned into fine grained optimizations that can't be replicated on PC

It's not quite that you can't replicate them on the PC, it's that the problematic state of having to make a game workable for not only multiple different GPUs with the same µArch, but also different vendor µArchs makes it basically impossible to get the same level of optimisation across all these disparate configurations.

Windows being a general use OS doesn't help either - no "game mode" can make up for the OS not being solely dedicated to gaming.

IMHO MS should have made a gaming only version of Winodws for multi boot years ago with everything but the essentials stripped out.

Now that is a feature I could have gotten behind.

 

[Mopetar]

If it's semi-custom and they're already cutting stuff, I'd cut any RT hardware that isn't used for DLSS and ask NVidia to backport their newer DLSS technology as that'll be a more valuable use of space as people will dock these with their 4K TVs and being able to upscale will be important.

Based on the linked article and Ampere designs, this would have 12 RT hardware accelerators and they'd be running at ~500 MHz. Utterly pointless to include them even if they were still targeting 720p.

No amount of magical R&D can fix reality. No companies are going to want to waste time trying to get ray tracing working on a console that people aren't buying for the graphics. Nintendo long ago figured out that they can use art direction to produce beautiful looking games that aren't pushing polygons, and they often hold up better after a decade when anything pushing the boundaries of photorealism winds up looking dated.

 

[Saylick]

If it's semi-custom and they're already cutting stuff, I'd cut any RT hardware that isn't used for DLSS and ask NVidia to backport their newer DLSS technology as that'll be a more valuable use of space as people will dock these with their 4K TVs and being able to upscale will be important.

Based on the linked article and Ampere designs, this would have 12 RT hardware accelerators and they'd be running at ~500 MHz. Utterly pointless to include them even if they were still targeting 720p.

No amount of magical R&D can fix reality. No companies are going to want to waste time trying to get ray tracing working on a console that people aren't buying for the graphics. Nintendo long ago figured out that they can use art direction to produce beautiful looking games that aren't pushing polygons, and they often hold up better after a decade when anything pushing the boundaries of photorealism winds up looking dated.

I'm sure Nintendo considered asking Nvidia to just either strip out or turn-off the RT hardware, but the fact that it's kept in likely means Nintendo wanted it in there. How much of a benefit will a handful of RT units running at low clocks will have is to-be-determined, but I understand your concern since there have been documented cases of Nvidia's technology becoming worthless when there simply isn't enough compute on the chip (see RTX 4060). The beefed up Optical Flow Accelerator seems to imply Nintendo will use some sort of Frame Generation technology to boost frame rates to alleviate the high compute requirements w/ RT.

 

[Ghostsonplanets]

It's not quite that you can't replicate them on the PC, it's that the problematic state of having to make a game workable for not only multiple different GPUs with the same µArch, but also different vendor µArchs makes it basically impossible to get the same level of optimisation across all these disparate configurations.

Exactly

Utterly pointless to include them even if they were still targeting 720p.

1080p screen Portable
4K output Docked

Nintendo long ago figured out that they can use art direction to produce beautiful looking games that aren't pushing polygons

While true, Nintendo games still make usage of modern rendering pipeline. For ToTK they even developed their own custom virtual geometry like solution. Or Monolith developing and shipping their own temporal upscaler solution. Countless examples, really.

The beefed up Optical Flow Accelerator seems to imply Nintendo will use some sort of Frame Generation technology to boost frame rates to alleviate the high compute requirements w/ RT.

I wouldn't be so hasty to make such judgment. While true that the OFA is beefed up compared to Ampere, I'm not sure if that's enough for DLSS Frame Generation given it's fairly expensive.

 

[Saylick]

I wouldn't be so hasty to make such judgment. While true that the OFA is beefed up compared to Ampere, I'm not sure if that's enough for DLSS Frame Generation given it's fairly expensive.

If not for video interpolation/extrapolation, what else would a beefed up OFA be used for? If it's not beefy enough for DLSS Frame Generation, why bother keeping it? For video playback?

 

[Mopetar]

Yes the benefit of consoles is having a simpler hardware target, but that only works up until you want to be cross platform at which point you're now targeting two or three hardware platforms and probably the PC anyway. Given those are all x86 platforms, Nintendo is already the odd one out by using ARM even though NVidia is more popular in PC graphics.

It doesn't matter if it's someone only targeting the Switch 2 for their game. They aren't going to add RT unless they're making a PowerPoint presentation. Unless it's also useful for some other common calculations they'd otherwise need to do on the shaders at a much greater cost, there's no point in having that hardware there.

I can see them wanting frame generation regardless of whether they use RT or not because it means they can use an SoC that only needs to hit 45 FPS to spit out 60 FPS. The DLSS upscaling is already obvious given they want a nicer looking 4K image when docked.

I don't think the people who are pushing this idea of RT realize how much it murders performance. Here's an image from a 3050 review showing RT results for Cyberpunk.

30 FPS for 1080p Medium settings. That's for a card with around 66% more cores operating at double or even triple the frequency. You'd need mythical Vega-level magic drivers just to get barely acceptable frame rates in a ~4 year old game at medium settings.

Maybe Sony or Microsoft would try chasing something like that, but this is Nintendo, a company that for almost two decades now has gone in the opposite direction of avoiding expensive graphical frills that add very little to the kind of games they want to make while incurring a large amount of expense.

We already have a significantly less powerful platform than other consoles and PCs that will be using a different CPU and GPU architecture from the other consoles. That's two strikes. The third is that the base for Nintendo are mainly there to buy Nintendo games. Unless they don't own a PC, why would they want to buy a third part game that aims at greater graphical realism when they could get that for one of their other consoles or their much more powerful PC instead? Strike three.

The only reason for Nintendo to include RT hardware is that NVidia asks them to or gives them a deal if they do have it. Even if it won't be used it probably lets NVidia add another however many million devices to their marketing material evangelizing their RT.

 

[Leeea]

I think Nintendo is planning to use it.

But not in high polygon games. It would be a great add on for simple puzzle games, and things like that.

Want tetris to look better quick and easy? 3d it, and then RT it. Its still tetris, it is still a 2d game, but it is set in a raytraced 3d world.

2D platformers, same thing.


My point is, there are lots of games out there that are no where close to using all the power. Those would be candidates for RT.

 

[poke01]

Nintendo doesn't go all out with textures, I think their dev teams can find creative uses for RT. If they can make a game like tears of the kingdom on such old hardware then Nintendo can surprise us.

 

[DeathReborn]

I think Nintendo is planning to use it.

But not in high polygon games. It would be a great add on for simple puzzle games, and things like that.

Want tetris to look better quick and easy? 3d it, and then RT it. Its still tetris, it is still a 2d game, but it is set in a raytraced 3d world.

2D platformers, same thing.


My point is, there are lots of games out there that are no where close to using all the power. Those would be candidates for RT.

Monkey Ball with RT reflections on the ball is what comes to my mind, maybe even using it just for lighting & shadows in caves/dungeons in Zelda or whatever.

 

[Mopetar]

Sounds like a lot of solutions in search of a problem.

My 2D platforming games really need to add a realistic lighting system for all of the shadows and reflections they have. My monkey's ball needs a realistic reflection that I can stare at as he goes over a ledge because the controls are unresponsive when the game is dropping to 24 FPS instead of running at 60+. I really must have an RT rerelease of Rayman since it's got "Ray" in the title, only they can call it RayTraceMan this time! Pretty clever!

Okay, hands up, how many of you bought Sega consoles because of blast processing?

 

[jpiniero]

Okay, hands up, how many of you bought Sega consoles because of blast processing?

Now that you mention it... using RT as a marketing point wouldn't be that surprising.

I think it will be in there regardless mainly because I don't think the chip is exclusive. Perhaps they will come to an agreement to use the busted chips on a Shield TV or something down the line.

 

[Ghostsonplanets]

There are literally examples of SteamDeck running titles with RT enabled at 30+ FPS. And that's a PC with RDNA 2 RT units. Even the Switch has a shipped commercial title with SVOGI enabled. But somehow a wider GPU with more performant RT units won't be able to leverage it?

 

[Mopetar]

Now that you mention it... using RT as a marketing point wouldn't be that surprising.

I think it will be in there regardless mainly because I don't think the chip is exclusive. Perhaps they will come to an agreement to use the busted chips on a Shield TV or something down the line.

It's doubtful that Nintendo would allow the severely busted chips to be used in competing products. Since it's Samsung 8nm they're using the yields should be good enough that most of the chips are good and they might bin the chips slightly just to allow for them to use almost every chip.

If they have some that are completely unfit for use, they will have some of the hardware disabled and be sold into some small overseas market just like the console SoCs for Sony / Microsoft.

There are literally examples of SteamDeck running titles with RT enabled at 30+ FPS. And that's a PC with RDNA 2 RT units. Even the Switch has a shipped commercial title with SVOGI enabled. But somehow a wider GPU with more performant RT units won't be able to leverage it?

I'm sure if you drop the settings and the resolution low enough you can get the game to run fast enough to barely manage to have the look of an old time movie. I'll bet the reflections of the ultra low quality textures look amazing though. Much better than the higher quality textures or the higher resolution or even having a higher frame rate.

This Eurogamer article about RT on the Steam Deck makes me feel so great about the prospect of RT on Switch 2.

We're mostly pegged to the low settings preset here with medium RT enabled. Control is running at just 540p internally, but gets temporally upscaled to 720p

Performance starts off strong - 30-40fps in the early game - but the game sees more complex environments as we continue, so performance drops into the 20s later on.

The default configuration is adept, with a 50-60fps readout at low settings, FSR 2 performance and RT disabled, but engaged RT flattens performance into the mid teens. Surprisingly, the full path tracing experience technically runs on the Steam Deck... though frame-rates are in slideshow territory at just 2-3fps.

Rounding out the SteamOS games today, we have Metro Exodus Enhanced Edition, where I've targeted low settings with medium RT - specifically RTGI. The key concession here is resolution, which I pegged to just 960x600 to boost frame-rates as much as possible. With those options the game largely runs above 30fps, but I'd imagine some later sections such as the demanding Taiga level would produce deeper lows.

Yeah, I'm gonna have to pass on that.

 

[soresu]

It's doubtful that Nintendo would allow the severely busted chips to be used in competing products.

It's not a question of what is allowed, but rather who owns the chip contract.

If it's nVidia and they just sell the chips on to Nintendo then of course they can do whatever they like with their own chips.

If OTOH it's Nintendo then ye it's not going to be used in anything else.

Though to call SHIELD TV a competing product to Switch is a serious stretch.

 

[poke01]

It's not a question of what is allowed, but rather who owns the chip contract.

If it's nVidia and they just sell the chips on to Nintendo then of course they can do whatever they like with their own chips.

If OTOH it's Nintendo then ye it's not going to be used in anything else.

Though to call SHIELD TV a competing product to Switch is a serious stretch.

If a new shield tv based on the T239 releases that would be very nice. Right now the Apple TV 4K A15 is the best streaming box. The Tegra X1 is too slow. Heck even my A10X Apple TV(still getting updates, unlike the shield) the is still smoother than the Tegra X1 Shield.

So hopefully Nvidia owns the contract

 

[Mopetar]

It's not a question of what is allowed, but rather who owns the chip contract.
. . .
Though to call SHIELD TV a competing product to Switch is a serious stretch.

I'm assuming based on what others have shared that Nintendo owns enough that NVidia isn't making a similar chip for anyone else.

With the Switch, Nintendo just decided to use NVidia's existing Tegra X1 since it generally fit their needs and NVidia needed someone to buy them since their other partners had pretty much given up on Tegra at that point. The only product other than Nintendo or Nvidia to use the X1 was a Google Chromebook. The Switch saved the Tegra X1 from irrelevancy.

A lot of those NVidia devices were probably more of a way to showcase the technology in a product than any real desire to own some particular market. It seems like their Tegra line has been aimed more to cater to automotive companies as it did find a niche there, so Nintendo can probably negotiate some terms with Nvidia, particularly if both companies expect another 100+ million in sales.

I'm also thinking that Nintendo will bin the chips a little bit, just to maximize their yields. Even though Samsung 8nm will be quite mature by now, Nintendo still has some strict performance and power targets because this is going into a handheld. Even though it's an ARM SoC, it'll still be fairly large with 8 cores and 12 SMs. I'd guess they'd disable one CPU core and one SM. Even if they all function, they can still disable the weakest one for better power/performance.

I'm also somewhat curious if they'd talk with Nvidia about putting a small NPU on the SoC. Having something that they could offload game AI to would be really interesting. That might be a bit too forward looking, but most games still have awful AI and some hardware aimed at accelerating those operations and some dev tools aimed at helping developers build/train an AI would be a selling point. Probably not something that Nintendo cares about, but I wouldn't be surprised if either Sony or Microsoft were to go down that route.

 

[Ghostsonplanets]

I'd guess they'd disable one CPU core and one SM. Even if they all function, they can still disable the weakest one for better power/performance.

There's no core disabled. NVN2 API match configuration is 12 SMs/6TPCs/1GPC and 8 Cores. If there’s anything disabled, it means the full die is even bigger than 12 SMs and 8 cores for them to be disabled.

I'm also somewhat curious if they'd talk with Nvidia about putting a small NPU on the SoC.

DLA is specifically one of the things that were discarded from T234 for T239. The only MatMul accelerator on T239 are the Tensor Cores.

Even though Samsung 8nm will be quite mature by now, Nintendo still has some strict performance and power targets because this is going into a handheld.

Correct. I'm actually hearings whispers of something that might mean either T239 is crazy efficient or Nintendo is using fairly low clocks. But can't confirm anything and it's best to wait until Nintendo unveil the system later this year.

 

[yottabit]

If the hardware supports it I’m confident Nintendo would make use of the RT cores somehow (for first party titles). Ever since the Gamecube their whole aesthetic has been based on low/mid poly with simple textures, but excellent stylized shaders.

Not all RT tasks murder performance, they could incorporate it into a dynamic lighting system for instance just throwing a few extra rays to just improve the accuracy of dynamic lights somewhat. I doubt you’ll be seeing any RT reflections outside of cutscenes at least.

I also doubt 3rd party titles would use it, unless Nintendo forces them to which they have had a habit of doing historically if it’s going to be one of the marketed gimmicks of the console. Then expect a lot of token RT implementations