Topic: The Nintendo Switch Thread

@Dezzy
It's nothing like mClassic.

It's deep learning super sampling. They feed a server low res images and ultra-high res images and it "learns" how to upscale the lower res images to look like the ultra-high res images. They do this with millions upon millions of images and train it. Each game has to be "trained" into the system separately.

Once complete, DLSS uses special AI accelerators called Tensor cores to process and upscale the rendered image in real time. But it does it before the signal is ever output. It has to be built in on the chip level.

Part of this is because in order to upscale the next frame it has to anticipate what the next frame will be, and it determines this by motion vectors embedded in the data.

Thus far, Tensor Cores are only in Nvidia RTX 20xx and 30xx GPUs (as far as gaming GPUs- I imagine they may be used in other product lines for automobiles). My RTX 2080 ti has them and can do raytracing and DLSS 2.0. The new GPUs (RTX 3070, 3080 and 3090, for example) can also do it. But if a mobile chip were to integrate Tensor Cores, it could do it too. But it's not something that can be done externally. It would require a new chip design with integrated Tensor Cores for the machine learning.

@JaxonH

Right, but all of that is still post-processing though, so I'm not sure why it couldn't be separated in principle. Maybe it hasn't in practice so far, but that's a different issue.

@Dezzy
It’s not “post processing” though (at least, not in the terms you’re speaking of- like mClassic taking an already rendered 2D image and just sprucing it up- traditional post processing must happen before the signal is output, unlike mCable).

It’s part of the rendering “process”- I think that’s a better way to put it. Another way to put it is to clarify the difference between external post processing and internal post processing. mClassic is external. It’s just taking a 2D image and trying to use anti aliasing to clean up the jagged edge lines. Traditional post processing, and DLSS especially, are done internally, before the image is sent out as a 2D signal. It’s why the motion vectors are needed so that it can predict the next frame and render that upscaled frame with the tensor cores predicting what it should look like. It’s why tensor cores are needed, built into the die of the chip itself.

Basically what I’m saying is, it can’t happen in the dock.

@JaxonH

How is it not? Motion vectors are in screen-space as well (at least every implementation I'm aware of, which is mostly Unreal and Unity), so that definitely counts as post processing. Every part of the algorithm that I'm aware of (which I only have a layman's understanding of, so maybe I'm missing something), happens in screen-space. That counts as post-processing by my lights.

@Dezzy
Again, read my above post.

There is a difference between external post processing an internal post processing. mClassic externally upscales a 2D image and cleans up jagged lines in a 2D image. Traditional post processing happens before the signal is output. It’s work done by the processor itself (and this explains why mClassic AA is far worse than traditional post processing).

DLSS requires Tensor cores built into the chip, and it must happen before the signal is output.

https://www.techspot.com/article/2049-what-are-tensor-cores/

I’m not up to speed on the cutting edge of external, auxiliary processing via USB-C connection- all I know is that the only GPUs that can do DLSS need Tensor Cores built in. Maybe that’ll change one day, but I’m not currently aware of anything that can do that as of right now.

@JaxonH

"upscales a 2D image and cleans up jagged lines in a 2D image" is exactly the same thing traditional post processing would be doing though, so I'm not sure what the principled distinction is.

I've read that article, and the only thing I could find that actually speaks to the question of why it needs to be built into the GPU, is that they plan to use tensor cores for other things as well, like physics calculations. That certainly would require it being part of the GPU, because it would need to be done in world-space.

@Dezzy
I think that’s where the misunderstanding is coming in.

Cleaning up a 2D image is not what traditional post processing does. It does it before it’s ever creates the 2-D image signal. There is a very distinct difference between mclassic and traditional post processing..

Per Eurogamer:

So with all the setup done, what does the cable actually do? We've heard claims that the effect is similar to SMAA setting, with Marseille using a proprietary technique it calls contextual anti-aliasing (CXAA). The comparisons do have merit, even if it's not precisely the same. SMAA can be integrated into the rendering pipeline, whereas the mCable only gets the final, flat, 2D output of the console to look at. The mCable can't provide a deeper, engine-level understanding of the game, or an intelligent use of previous frames similar to temporal anti-aliasing. What you have instead is all done in post

The Tensor cores are required for the Tensor calculations, which GPUs are very good at. It takes a whole lot of CPU cores to do the same matrix multiplication operations a GPU can do. It also takes a lot of power and so they pack in dedicated cores specifically devoted to that.

I mean, I’m not sure what this is even about, or why you’re so insistent DLSS can be done from the dock without Tensor Cores? Even if it could be done, let’s just say, it’s not something that’s likely going to be done, so why even bother discussing it? There’s no clear advantage.

It’s like, trying to reinvent the wheel when the wheel’s already been invented. Use a chip with integrated tensor cores and voilà, you’ve got DLSS capabilities, with no drawbacks, for both docked and handheld. Do you think they’re going to gather their engineers and try to figure out a way to send a signal to the dock, deconstruct the image (if that’s even possible, I don’t know), let a built in dock GPU run all the calculations, recompile the image and send it back out (hopefully avoiding any lag in the process) which still leaves handheld mode lacking any benefit, rather than just... using a chip with Tensor cores? To what end?

I’m all for exploring different possibilities, but, there’s no evidence such a thing could even exist, and even if it could exist, there’s no evidence that it provides any benefit, and even if it could exist and could provide benefit, there’s no evidence it’s worth investing the time or money to figure out how, and there’s still the glaring issue of handheld not gaining any benefit, so any game being played that requires that power, when you undock the system, would be unplayable, crash the system, or worse. So... the whole conversation kinda seems moot.

I’m not saying an auxiliary processing unit in the dock is ruled out, but I think it’s fair to say if they take that particular approach, DLSS is almost certainly off the table. And if such an approach was ever to be utilized, it would all but ensure the additional power doesn’t get taken advantage with games that couldn’t otherwise run on the system. For the same reason Nintendo didn’t let us use external hard drives (undocking would crash the game) I don’t think they would be in favor of auxiliary power. Unless it was negligible to the point it couldn’t make new games possible on the system, and only enhanced playability of existing games.

@JaxonH @Dezzy
I see no reason why they couldn't in theory add post processing to the dock. The problem is for it to be worthwhile and better than the kind of upscaling your TV could do it would increase the cost of the dock significantly. To the point where they'd be losing money on the standalone dock which is significant because I'm fairly confident there's a significant markup on the standalone dock.

But for less than that they could release a new SKU with more horsepower in general, decent DLSS upscaling built in and include a higher HDMI spec dock. Which would give performance advantages both docked and undocked AND give the Switch a bit more momentum at some point next year.

@JaxonH

I see your misunderstanding and I raise you!

The confusion is that you're conflating anti-aliasing specifically with post-processing in general. There are like 20 different mainstream anti-aliasing approaches. Some of them work as regular post-processing shaders (e.g FXAA, NFAA), but others don't, or are just automated by the GPU hardware (as MSAA usually is). There's even one that tries to run on the CPU, which sounds kind of mental to me, but it exists all the same.

So when I said MClassic works the same as a regular post-processing shader, I assumed we meant if it was using the same algorithm. FXAA implemented by an MClassic would be exactly the same as FXAA done by the GPU, so that's something you could separate into a different piece of hardware with no problems.

The general point I was trying to get at, was that it might be easier and cheaper for them to develop if they could separate that DLSS function into the dock, because then they would have to worry less about how to get the necessary power source and cooling system into a handheld device (that they want to fit into someone's pocket)

@skywake
Post processing, sure. I’ve never disputed the fact they could, in theory, use auxiliary processing in the dock. And certainly it could be better than upscaling your TV does.

But DLSS specifically? That’s something else entirely. DLSS is not just fancy upscaling. It’s a taxing process utilizing cores devoted to advanced machine learning prediction within the rendering pipeline. If it was that easy then we would see auxiliary devices offered to make DLSS viable on PCs that don’t have RTX 20xx and 30xx GPUs with Tensor Cores (the vast majority of the PC market).

That said, there is no real reason to think they’d go for auxiliary power in the dock regardless. Even if we weren’t hearing about the DLSS (which pretty much writes it off entirely, since you need a chip with Tensor Cores), the marginal benefit you get from additional processing in the dock could simply be done with a more powerful chip that benefits both docked and handheld mode.

The only way we see auxiliary dock processing is if it’s negligible increase. Otherwise why bother. If you’re going to make it a large increase then you want to get bang for your buck for that investment. Such as games running that couldn’t previously run. And if you do that when it’s only possible in the dock, well, the problem is self evident. You undock, and the game crashes.

So if DLSS is true, it’s a new chip using Volta or Ampere with Tensor Cores. Open and shut. If DLSS is not true, I still don’t think they’re going to do auxiliary power in the dock for the reasons stated above (although I won’t go so far as discounting it entirely). Why not just make a more powerful chip that benefits both modes of use, which is basically what you’re saying.

Right. My assumption was that the cost of the dock would just be substracted from the cost of the console though, given that the equivalent tech would need to be added to the GPU otherwise. Probably not true in practice I realize, but in principle $A + $B= $(A+B)

@Dezzy
I’ll agree, it does seem it would be easier to implement (not necessarily cheaper but maybe) versus changing the chip IF they could figure out a way to separate into the dock. But that’s something I’ve seen no evidence of being possible thus far. But something to keep in mind is, DLSS cannot be grouped into other forms of post processing such as anti-aliasing, because it’s not the same, even if it is technically a form of processing that occurs post-initial rendering. And even anti-aliasing that occurs within the rendering pipeline, they have not figured out how to do externally (hence mClassic using AA on the flat, 2D image signal instead), so if they can’t even figure out how to do that, then I see no reason to believe they could figure out how to do it for something infinitely more complex that requires it’s own Tensor Cores on the chip, and would ultimately introduce a fatal design flaw of games not working when undocked. So perhaps it’s scientifically possible to create such an architecture if they put their minds to it (I’m sure anything is technically possible- the time, cost and pros/cons is another discussion entirely) but I don’t see why they would. That seems like a massive undertaking for design engineers when they’ve got a solution already present.

The thing is, using a chip with Tensor cores would inherently mean using a much, MUCH more efficient chip that draws MUCH less power. We saw the gains from dropping down to 16 nm from 20 nm. A 50% reduction in power draw. A drop to 8 nm (or even smaller), along with a screen that draws less power (which we know they’re looking into, and that’s something they wouldn’t be doing if they didn’t need the gains in handheld mode), could mean hitting 8th gen power.

Currently, Switch handheld runs on 15W max. But the v2 model reduced average use down into single digits. If they released a new model that used all 15W or even bumped it to 20W, with a bigger battery, way more efficient screen, WAY more efficient chip... I think it’s possible.

Of course, handheld mode would be rendering in 720p at most which even the current Switch can do. DLSS would ensure it can output a 1080p image at the rendering cost of 720p. Couple that with a more efficient screen, more efficient chip and potentially better battery, and we could have 1080p 60 FPS with 5+ hrs battery life minimum.

The only hang up I have, is backward compatibility. DLSS means Tensor Cores. And Tensor Cores means Volta or Ampere architecture. And Volta or Ampere architecture means a significant change from the current X1 design. Is it possible to switch to such a radically different chip design and maintain full compatibility, both forward and backward? It would have to, if it’s a true revision.

I think times are changing, and I think the old concept of 7 year generations with a clean break to new hardware is outdated. I think third parties are reluctant investing in an ecosystem that has to be built from scratch every 7 years. I think third parties want dependable and predictable returns on their investments, and the old system is just not conducive to such goals. The uncertainty from the up/down cycles of Nintendo consoles in years past is not sustainable enough to entice 3rd parties long term. And I think all of this points to a new paradigm that more closely resembles the mobile market than traditional console generations. The fact so many games are just being released as PS4/X1 games that “can also run on PS5/XS” instead of as pure PS5/XS games, is evidence of 3rd parties’ reluctance to start from scratch, even with their most reliable markets.

Moving forward, I think new generations will be integrated as revisions with full BC, which are more frequent than traditional console cycles (such as 5 year refreshes). Mobile tech is advancing at a higher rate of increase than PC tech. So while traditional consoles may be able to last 7 years, Mobile consoles become outdated far quicker, and thus need to refresh at a faster rate. This isn’t to say Nintendo couldn’t come in with a brand new Switch successor later, in 2025 or 2026. But if they do, I suspect it won’t look much different than the revision, with one key difference: some new gimmick that differentiates it from the last generation which is something we know they value. But I still would expected to offer full BC. The only thing that throws a wrench into this idea is that it would still have the disadvantage of there being a cut off point where all games developed no longer work for the previous install base, and that is a disincentive to third-party developers having to start over. So perhaps we won’t see a next generation switch, at least, not in the way we expect. Perhaps it will just be revisions, every 5 years. And maybe they call the next one the next generation Switch, but many games can still run on the original Switch. Or maybe, by that point games only work for one revision prior, meaning Switch Pro and Switch 2. Then Switch 2 Pro games works for S2P and S2, then S3 games work on S3 and S2P.

Who knows. What I do know is, Nintendo needs to retain 3rd parties to remain relevant (and they know this too, hence them approaching so many big devs before launch) and a stable, predictable market is the key to pulling support long term. I’m sure they want a stable, predictable market for their own devs as well. Continuing to flip a coin and holding your breath to find out if the next generation is going to be a success isn’t a viable long-term strategy. They have found a concept that works and fills a niche nobody else offers. This is their chance to capitalize and lock it down. And you can’t lock it down without providing a stable, predictable market that is not susceptible to the ups and downs of past Nintendo console trends. And the only way to do that is to stop building a new install base over from scratch every 7 years. It needs to be a perpetual, evolving platform like iOS, where new revisions release, and the games for them can still run on the previous revision if power permits.

Well I assume they're not gonna have that function working when it's undocked either way. I assume the power draw will just be too high to run off a battery. Plus you don't really need any higher than 720p on a handheld screen. So either way, I'm assuming they're gonna be developing a version where the tensor cores only activate when docked.

@Magician we need to remember/consider that the Legend of Zelda franchise is notorious for the delays until the final release of it games, the significant impact of the pandemic in the development of Zelda Breath of the Wild 2 and the potencial release of the remaster of Wind Waker e Twilight Princess on Switch.

@Dezzy
I don’t think that’s the case. DLSS would make it less demanding to run in handheld, not more.

Handheld games could render at 1080p for the cost of rendering 540p or 720p. That’s the whole benefit. It lowers the power required for the same performance, or increases the performance for the same power drawn.

So ya, DLSS in handheld would mean getting 1080p on every single game across the board, with battery life equivalent to rendering 540p or 720p.

I’m really enjoying Persona 5 Strikers

Definitely a little pixelated on Switch but once I adjusted it was fine. The sprites are perfect, closer shots of 3D models looks good, audio is fantastic, music is excellent, gameplay feels VERY similar to Persona 5. It really does feel like a Persona experience. It’s not exactly the same. Action combat for example. But all the same characters with the same voice actors with the same writing and a heavy focus on story, with a lot of similar mechanics like All Out Attacks, Persona skills, items, equipping gear, accessories, etc. You explore jails which are just like palace dungeons, there’s the Velvet Room. The menus have the same snazzy animations and transition screens still show the calendar changing...

It feels like Persona. It really does. I’m having just as much fun playing this as when I was playing Persona 5. Highly recommended to anyone looking for a Persona experience. It almost doesn’t even feel like a spin off. Like, Persona Q, Persona Dancing, Persona Arena... these felt like spin-offs. And I still enjoyed them- I played Persona 4 Dancing as my first Persona game, and it’s the only game I’ve ever platinumed. Those still integrated the Persona experience, at least to an extent, into an entirely separate genre. This feels like the same genre, but with action combat and some slight differences. It feels 80% the same. I really do love it.

Hi everyone!

@JaxonH It's insane how close they got to emulating the structure of mainline Persona minus some key elements that I can actually excuse them not being there for the sake of streamlining the story to fit the gameplay

Although those opening hours, I can see people who haven't necessarily at least read up on Persona 5's plot being a bit lost as to who these characters are because the game doesn't necessarily care about catching up newcomers. I'm already seeing references to stuff like "last year's Change of Heart incident", the Meta-Nav app, Haru's love for gardening and Yusuke's love for art that I think might fly over some people's heads, but it's standalone enough where I think for the most part in terms of the story specific to this game they can come into it and be mostly fine. The little things are still there for people who've played P5 or P5R though and it's nice attention to detail. They especially nailed the look of Yongen-jaya from the OG game

Also lmao your first Persona game was P4D? That's a first I've seen

I think the Switch is good into early 2022 if they do absolutely nothing new for the hardware. After that point... indies and 1st-party stuff as far as the eye can see lol. Which would be wicked sad. I absolutely LOVED playing Spyro on my Switch, and the new DOOM games, and I'm looking forward to Crash 4. It may not be much, but those were purchases I would have made on other consoles/PC if the Switch versions didn't exist.

Monster Hunter Rise is a perfect example of how Nintendo's stance on 3rd parties has changed. Monster Hunter was a big franchise on 3DS but there were big 1st party releases nearby. For example:

• Monster Hunter 4 Ultimate - Majora's Mask 3D launched same day
• Monster Hunter 3 Ultimate - Luigi's Mansion 2 launched less than a week later

Meanwhile 3rd party games like MH Rise pretty much have several months to themselves as the next big 1st party release after 3D World last week is Mario Golf Super Rush on June 25th.