StarTopic Future Nintendo Hardware & Technology Speculation & Discussion |ST| (New Staff Post, Please read)

[Paul_Subsonic]

for shadows, NRD has SIGMA denoiser. that said, ReBLUR is used for WatchDogs Legion I think, though I guess the low RT resolution allows it to overcome costs

Actually, the denoiser for WatchDogs is different on consoles. It's ReBLUR on PC, but Digital Foundry's video shows the console denoising is vastly inferior.

Once again showing that even on the big boxes, ReBLUR is no easy feat.

 

[ILikeFeet]

Oops I mistook ReLAX with SIGMA

I have reserached the subject of the cost of denoisers for hours yet I make this mistake. My stupidity never fails to surprise me.

I found the costs of ReBLUR

it makes ReLAX more expensive, so most games probably won't use NRD to begin with. I wonder how SIGMA fairs since that's just for shadows. and TAA is also an option, so maybe DLSS can work as a natural denoiser

 

[7hatDeadCat]

2048 CUDA Cores@750MHz = 1536 CUDA Cores@1GHz

3 TFLOPs = 3 TFLOPs

that might no be entirely true as some aspects of the gpu may have different performance benefits from being wider, or faster.

 

[Ghostsonplanets]

that might no be entirely true as some aspects of the gpu may have different performance benefits from being wider, or faster.

Yes. But the theoretical FP performance is the same and is the most apples to apples comparison that can be made.

 

[Paul_Subsonic]

I found the costs of ReBLUR

it makes ReLAX more expensive, so most games probably won't use NRD to begin with

This is Cyberpunk IIRC ? It's from the "Ray Tracing Gems" book, that's for sure.

I did take a look at it a few weeks ago (it's free and constains so much info), and it also shows Control's denoising, which is much more realistic for Drake.
For a limited number of effects, that is; the multiplication of RT effects makes Control approach actually perform worse.

 

[Discostew]

Something I've been wondering. GPUs love high bandwidth for rendering, but wasn't it said that RT also benefits from low-latency RAM? What about DLSS? Even Rich's example with the RTX 2050 still uses high-latency GDDR for VRAM. I can't even recall if there has been an RTX card that hasn't used GDDR.

 

[ItWasMeantToBe19]

I found the costs of ReBLUR

it makes ReLAX more expensive, so most games probably won't use NRD to begin with. I wonder how SIGMA fairs since that's just for shadows. and TAA is also an option, so maybe DLSS can work as a natural denoiser

I posted this like a week ago

 

[ILikeFeet]

This is Cyberpunk IIRC ? It's from the "Ray Tracing Gems" book, that's for sure.

I did take a look at it a few weeks ago (it's free and constains so much info), and it also shows Control's denoising, which is much more realistic for Drake.
For a limited number of effects, that is; the multiplication of RT effects makes Control approach actually perform worse.

aye, it's from there. I don't know if it's from Cyberpunk specifically. the section doesn't specify and seems to cite CP2077 for specific figures.

my personal expectation is that games on Drake will largely use one or two RT effects, so ray reconstruction's benefits might not be valued so much.

 

[Ghostsonplanets]

Something I've been wondering. GPUs love high bandwidth for rendering, but wasn't it said that RT also benefits from low-latency RAM? What about DLSS? Even Rich's example with the RTX 2050 still uses high-latency GDDR for VRAM. I can't even recall if there has been an RTX card that hasn't used GDDR.

GPUs use GDDR because, as you said, they love bandwidth. Even if RT like low latency, LPDDR is too slow to provide the bandwidth needed by GPUs in the same area as GDDR. The reason Phones, Laptops and portable consoles like Switch use LPDDR is because energy-efficiency is key for these devices and they trade off bandwidth for energy-efficiency.

 

[Paul_Subsonic]

Maybe I should just dump every shred of info on the cost of denoisers I found.
Maybe I'll get around to finish that damn breakdown and calculator, but in the mean time lots of people here might be interested in the raw data.

 

[Paul_Subsonic]

I posted this like a week ago

As they say in my country :

Aïe. Coup dur pour Guillaume.

 

[BreakAtmo]

2048 CUDA Cores@750MHz = 1536 CUDA Cores@1GHz

3 TFLOPs = 3 TFLOPs

I remember that in the Road to PS5 presentation, Mark Cerny used basically this exact comparison to talk about how not all TFs are created equal. When you increase GPU clock speeds, you also boost the other parts of the GPU that aren't just the SMs/CUs. Those parts don't benefit in the same way from there being more SM/CUs.

 

[Deleted member 887]

A (long) Note About The DF Video and How To Interpret The Results
I want to quote a DF Direct real quick (edited for clarity)

Rich: Bobby Kotick is saying that basically we're looking at last gen level performance, you know PS4 style. Which I think would be you know not bad for a mobile device but the point is I think that is only a fraction of what these new machines actually offerwouldn't you agree Oliver?

Oliver: Yeah I think it's an interesting quote that nonetheless doesn't totally align with what we think we know about the Switch 2 too because, just as one example, could a PS4 play a game like breath of the wild at an effective 4K 60 Resolution right which is the the rumor?[ ...] So in terms of the raw Hardware horsepower I would suspect the switch 2 could be falling in line with the capabilities of a PlayStation 4 or an Xbox one but uh I don't really think in terms of the actual rendering power the switch 2 might might bring to the table in terms of offering dlss uh bringing that to the table I think the switch 2 should should be able to outperform those consoles at least in scenarios where the technology is being effectively used

Consoles are weird machines. CPUs and GPUs both have new major release every two years, and the industry is constantly pushing those boundaries. A console needs to provide a great bang for the buck now, while making decisions that will allow the technology to hold up 7 years from now.

This is the way to think about Rich's tests. Nvidia brings technologies to the table that have, even in their budget products, never been tested in such a small device. It brings up a number of questions, like -

• How good does a console implementation of DLSS - lower than 4k res, on a big TV - look?
• How well does DLSS perform on a low power device?
• How good is Nvidia's Ray Tracing hardware on a lower power device?
• What could such a small GPU do, if paired with a more modern CPU/Storage solution?

Let's look at the games and see what lessons we can extract instead of "this looks good" or "that runs poorly" or "I don't like that resolution."

Death Stranding: PS4 delivered 1080p30 on this game, and Rich shows that native. PS4 Pro delivered 4k checkerboarded, also 30fps, the test machine does it with DLSS Ultra Performance mode, though somewhat unstably. Rich also shows a 1440p30 mode that is much more stable.

What did we learn: The drum I have beat is "PS4 power, PS4 Pro experiences possible, but different tech means devs might make different decisions." Here we see all of that hold up. The raw power is plenty good enough to just "do" the PS4 experience without any real work. A 4k DLSS experience hits some performance snags that would be ironed out by a quality optimized port.

We also see the DLSS is a totally different technology from checkerboarding. It looks better, it's more flexible, but its cost grows differently. This creates different tradeoffs, and we shouldn't expect devs to make exactly the same choices.

Cyberpunk 2077: Death Stranding is a last gen console game with a good PC port. Cyberpunk is a PC game with a shitty last gen console port. Rich shows us the game running at PS5 quality settings, but at 1080p30, with instability.

What did we learn: We start to see how and why developers might make different decisions than on the AMD consoles. PS4 Pro runs at 1080p30, with a series of settings that are described by DF themselves as "extremely blurry and just kind of visually kind of glitchy". Series S has a 1080p60 mode. Both are 4 TFLOP machines. Here we see the 3 TFLOP Ampere card absolutely smack the pant off the PS4 Pro, running at comparable frame rates and resolutions, with substantially higher quality settings, but unable to deliver the Series S 1080p60, even with DLSS enabled.

DLSS is a different tech, it doesn't behave like "just more flops."

A Plague Tale: Requiem: A game that didn't come to last generation consoles, that runs at 900p30fps on the Series S, is here comfortably at 1080p30fps.

What did we learn: That the GPU isn't all that matters. Plague Tale is rough on the GPU, sure, but it's famously CPU limited. Pairing even this weak GPU with a modern CPU and storage solution, and suddenly 9th gen exclusives become viable.

Control: This runs at 30fps on the last gen machines, and kinda badly at that. 1080p on the PS4 Pro. Here it runs more stably, but same resolution, matching the PS5's settings.

What did we learn: Here, once again, we have the PS4 Pro performance/resolution experience via DLSS, but that's not the interesting story. What's interesting is that we're getting that level of performance with ray tracing enabled. Compare to the PS5 - these settings are matched, with PS5 at 1440p30fps. The Series S can't deliver an RT mode at all. But here we actually have a case where something "comparable" to the PS5's RT mode is actually easier to achieve than an ugly-but-fast 60fps mode.

Again, just like DLSS, RT cores change the math, opening up options that aren't possible on other hardware. RT is viable.

Fortnite: Rich tests with Nanite on, Lumen with hardware RT, Virtual shadow maps at high. With DLSS, he gets a comfortable 1080p30.

What did we learn: "Do I wanna play Fornite at 30fps???" I dunno man, I don't know your life. Who cares, that's not what this is about. The next-gen game engine, running with it's entire feature set fully enabled, and is still delivering HD resolutions and acceptable frame rates.

This is the power of a modern device. At nearly every level, Nvidia is providing a more advanced solution than AMD. UE5 is built for temporal upscaling, and DLSS is best-of-breed. Nanite uses mesh shaders on hardware that supports it, and Ampere does. Lumen has a software fallback for better performance and older machines, but Nvidia's RT hardware performs nearly identically to the software solution.

"Is the Series S holding gaming back" is a dumb discussion point of the last few years. Now we get to ask "Is the PS5 holding gaming back?" With it's lack of mesh shaders, it's lack of decent hardware RT, it's lack of AI accelerated rendering - it's Nintendo that is making the promise of UE5 fully possible. And that's what Rich is demonstrating here.

 

[ILikeFeet]

I remember that in the Road to PS5 presentation, Mark Cerny used basically this exact comparison to talk about how not all TFs are created equal. When you increase GPU clock speeds, you also boost the other parts of the GPU that aren't just the SMs/CUs. Those parts don't benefit in the same way from there being more SM/CUs.

in the case, the TFs are equal. the design of the T239 is that of the desktop variant and the 2050. but the memory amount and bandwidth is definitely doing something to the results. too high for handheld and too low for docked

 

[Paul_Subsonic]

Part 1 of me dumping everything I saved on my phone and pc regarding denoising :

(Lots of it may be useless I'm literally just dumping everything)

Spoiler: stuff

, , , , , , , , , , , , , , , , , , , , , , , , , , , , ,

 

[SoldierDelta]

A (long) Note About The DF Video and How To Interpret The Results
I want to quote a DF Direct real quick (edited for clarity)



Consoles are weird machines. CPUs and GPUs both have new major release every two years, and the industry is constantly pushing those boundaries. A console needs to provide a great bang for the buck now, while making decisions that will allow the technology to hold up 7 years from now.

This is the way to think about Rich's tests. Nvidia brings technologies to the table that have, even in their budget products, never been tested in such a small device. It brings up a number of questions, like -

• How good does a console implementation of DLSS - lower than 4k res, on a big TV - look?
• How well does DLSS perform on a low power device?
• How good is Nvidia's Ray Tracing hardware on a lower power device?
• What could such a small GPU do, if paired with a more modern CPU/Storage solution?

Let's look at the games and see what lessons we can extract instead of "this looks good" or "that runs poorly" or "I don't like that resolution."

Death Stranding: PS4 delivered 1080p30 on this game, and Rich shows that native. PS4 Pro delivered 4k checkerboarded, also 30fps, the test machine does it with DLSS Ultra Performance mode, though somewhat unstably. Rich also shows a 1440p30 mode that is much more stable.

What did we learn: The drum I have beat is "PS4 power, PS4 Pro experiences possible, but different tech means devs might make different decisions." Here we see all of that hold up. The raw power is plenty good enough to just "do" the PS4 experience without any real work. A 4k DLSS experience hits some performance snags that would be ironed out by a quality optimized port.

We also see the DLSS is a totally different technology from checkerboarding. It looks better, it's more flexible, but its cost grows differently. This creates different tradeoffs, and we shouldn't expect devs to make exactly the same choices.

Cyberpunk 2077: Death Stranding is a last gen console game with a good PC port. Cyberpunk is a PC game with a shitty last gen console port. Rich shows us the game running at PS5 quality settings, but at 1080p30, with instability.

What did we learn: We start to see how and why developers might make different decisions than on the AMD consoles. PS4 Pro runs at 1080p30, with a series of settings that are described by DF themselves as "extremely blurry and just kind of visually kind of glitchy". Series S has a 1080p60 mode. Both are 4 TFLOP machines. Here we see the 3 TFLOP Ampere card absolutely smack the pant off the PS4 Pro, running at comparable frame rates and resolutions, with substantially higher quality settings, but unable to deliver the Series S 1080p60, even with DLSS enabled.

DLSS is a different tech, it doesn't behave like "just more flops."

A Plague Tale: Requiem: A game that didn't come to last generation consoles, that runs at 900p30fps on the Series S, is here comfortably at 1080p30fps.

What did we learn: That the GPU isn't all that matters. Plague Tale is rough on the GPU, sure, but it's famously CPU limited. Pairing even this weak GPU with a modern CPU and storage solution, and suddenly 9th gen exclusives become viable.

Control: This runs at 30fps on the last gen machines, and kinda badly at that. 1080p on the PS4 Pro. Here it runs more stably, but same resolution, matching the PS5's settings.

What did we learn: Here, once again, we have the PS4 Pro performance/resolution experience via DLSS, but that's not the interesting story. What's interesting is that we're getting that level of performance with ray tracing enabled. Compare to the PS5 - these settings are matched, with PS5 at 1440p30fps. The Series S can't deliver an RT mode at all. But here we actually have a case where something "comparable" to the PS5's RT mode is actually easier to achieve than an ugly-but-fast 60fps mode.

Again, just like DLSS, RT cores change the math, opening up options that aren't possible on other hardware. RT is viable.

Fortnite: Rich tests with Nanite on, Lumen with hardware RT, Virtual shadow maps at high. With DLSS, he gets a comfortable 1080p30.

What did we learn: "Do I wanna play Fornite at 30fps???" I dunno man, I don't know your life. Who cares, that's not what this is about. The next-gen game engine, running with it's entire feature set fully enabled, and is still delivering HD resolutions and acceptable frame rates.

This is the power of a modern device. At nearly every level, Nvidia is providing a more advanced solution than AMD. UE5 is built for temporal upscaling, and DLSS is best-of-breed. Nanite uses mesh shaders on hardware that supports it, and Ampere does. Lumen has a software fallback for better performance and older machines, but Nvidia's RT hardware performs nearly identically to the software solution.

"Is the Series S holding gaming back" is a dumb discussion point of the last few years. Now we get to ask "Is the PS5 holding gaming back?" With it's lack of mesh shaders, it's lack of decent hardware RT, it's lack of AI accelerated rendering - it's Nintendo that is making the promise of UE5 fully possible. And that's what Rich is demonstrating here.

If you don't get at least slightly excited reading all of this, we're not friends.

Special mention to this 11/10 line:

"Is the Series S holding gaming back" is a dumb discussion point of the last few years. Now we get to ask "Is the PS5 holding gaming back?"

Fucking hot knife through butter.

 

[CeramicPigeon]

I've been reading more and more analysis and I'm still kinda mixed.
I guess I retract what I said about it not being much better than the Steam Deck. but with the Gamescon thing and other prominent leakers backing up the 4K Zelda thing and DF saying that 4K DLSS realistically wouldn't be feasible makes me so split.

I'm sorry if this sounds stupid but it'll be 2024 when this thing comes out. It needs to be able to hit 4K. Maybe not consistently, as not even PS5/XSX can, but it needs to at least be the ceiling. Something that 2D games could easily hit, or maybe remasters of older titles from 6th or 7th gen.

 

[Paul_Subsonic]

Part 2

Spoiler: stuff 2 : the return

, , , , , , , , , , , , , , , , , , , , , , , ,
, ,

Those are all PDFs you can find online, the screenshots are only so you can see the title and find them. The screenshots themselves don't contain any interesting info.

 

[Samuspet]

Does anyone have a good theory on what could account for the huge disparity between DF testing, and the numbers in the DLSS documentation?

Real world performance in games and numbers on a spreadsheet are always different. It’s why teraflops are largely irrelevant and why you see PS5 at times outperform Xbox Series X despite it’s GPU being ~20% weaker on paper.

Games made for Switch 2 will be targeting it’s specific locked specs and Nvidia probably have a customised version DLSS baked into their Switch 2 API. Nintendo didn’t request tensor cores for shits and giggles.

 

[Ajimi]

Hi y'all, lots of interesting info lately

Just a very small update on the amiibo side ; I previously posted pictures of a new back packaging. Well, turns out we receiveid the new TotK Ganondorf and Zelda amiibos, and they still have the old back :

… Which is a bit weird to me, since I would expect brand-new amiibos to be designed with the new "universal" back. Oh well. Not the most interesting information of the past week, but I thought I would share anyway

Cheers !

Again a small update.

I didn't notice it before, but the LINK TotK amiibo has the "new" back (aka the neutral one, with no images in my previous previous post). So it's not just the Smash ones, and I do not know what logic there is behind the Ganondorf and Zelda ones. Wtf Nintendo.

 

[Hermii]

I've been reading more and more analysis and I'm still kinda mixed.
I guess I retract what I said about it not being much better than the Steam Deck. but with the Gamescon thing and other prominent leakers backing up the 4K Zelda thing and DF saying that 4K DLSS realistically wouldn't be feasible makes me so split.

I'm sorry if this sounds stupid but it'll be 2024 when this thing comes out. It needs to be able to hit 4K. Maybe not consistently, as not even PS5/XSX can, but it needs to at least be the ceiling. Something that 2D games could easily hit, or maybe remasters of older titles from 6th or 7th gen.

One thing about DLSS, is that it scales on output resolution. So scaling 3,99k to 4k would take exactly as long as scaling 720p to 4k. So yea, this machine probably isn't good enough for 4k dlss. It's probably able to run some games natively though.

My guess is the Zelda demo probably wasn't running at 4k, rumours like that should always be taken with a grain of salt anyway. Like fishing stories.

However I wouldn't be surprised if Zelda was able to hit 4k natively but 4k 60 probably not.

 

[darthdiablo]

Again a small update.

I didn't notice it before, but the LINK TotK amiibo has the "new" back (aka the neutral one, with no images in my previous previous post). So it's not just the Smash ones, and I do not know what logic there is behind the Ganondorf and Zelda ones. Wtf Nintendo.

Might be the timing of the transition from old to new backs.. Link happened to be made last in maufacturing after Ganondorf and Zelda, thus got the new back?

Odd division indeed (Ganondorf/Zelda having old back, Link having new one)

 

[Ghostsonplanets]

I'm sorry if this sounds stupid but it'll be 2024 when this thing comes out. It needs to be able to hit 4K. Maybe not consistently, as not even PS5/XSX can, but it needs to at least be the ceiling. Something that 2D games could easily hit, or maybe remasters of older titles from 6th or 7th gen.

I'm sorry but either you're not understanding what everyone is trying to tell you or just dooming for the sake of dooming. Not even PS5/Xbox Series will be able to hit 4K constantly. And we're already seeing both machines having to recosntruct from lower resolutions. Why should Switch 2 be bound to this? As everyone already is tired of saying: Switch 2 will be a powerful machine that will have GPU feature parity with the others consoles and will have comparable performance. Not even that, why should Switch 2 be doomed when Series S exists as this generation baseline?

Regardless, Switch 2 will be fine as long as it sell well. Just like with Switch, developers will try to do the impossible to fit their games into the machine, except that with Switch 2, they will be able to port and make their games run with much ease and without having to make significant downgrades.

I remember that in the Road to PS5 presentation, Mark Cerny used basically this exact comparison to talk about how not all TFs are created equal. When you increase GPU clock speeds, you also boost the other parts of the GPU that aren't just the SMs/CUs. Those parts don't benefit in the same way from there being more SM/CUs.

Yes, what you said is true. However, these are basically the same GPU IP. It's the best apples to apples comparison we can do.

 

[ILikeFeet]

one more example of denoising costs, this time from Qualcomm's UE5 talk on ray traced shadows. their lumen talk had nothing on denoising costs

 

[Molybdenum]

I thought I'd do a little math with the DLSS frametime cost from the DF video. First, the assumption made in the video is that the docked clock would be 1GHz, but we don't know for sure what the final clocks will be, and 1GHz is slower than a lot of the estimates from this thread. Assuming the DLSS performance scales linearly with GPU clock speed, then running the GPU faster would allow DLSS to take less time. Another common assumption we've been making is that docked clocks will be around double the portable clocks, and going by the portable clock estimate in the video of 540MHz, we would expect docked to be around 1080MHz, so slightly faster than what was tested in the video. Also, the switch's lowest portable speed was 40% of the docked speed, so I decided to check that one too. Also I'm assuming DLSS will be rendered concurrently with the next frame.

Here are a few clock speed and DLSS frametime combinations I've calculated:

• DF video: 1000MHz clock, 18.3ms DLSS (4k30 viable, 4k60 impossible)
• 2x DF portable clock estimate: 1080MHz clock, 16.94ms DLSS (4k60 just barely impossible)
• Minimum clock speed for 16.7ms DLSS: 1095.8MHz
• Docked clock estimate from earlier in this thread: 1100MHz, 16.63ms DLSS (4k60 just barely possible)
• Docked clock 2.5x estimated portable clock (like switch): 1350MHz, 13.56ms DLSS (plenty of room for 4k60)

In conclusion; DLSS to 4k60 is impossible if the docked clock is 1GHz like the DF video assumed, but it doesn't take much more docked speed to make it theoretically viable.

 

[Hermii]

I thought I'd do a little math with the DLSS frametime cost from the DF video. First, the assumption made in the video is that the docked clock would be 1GHz, but we don't know for sure what the final clocks will be, and 1GHz is slower than a lot of the estimates from this thread. Assuming the DLSS performance scales linearly with GPU clock speed, then running the GPU faster would allow DLSS to take less time. Another common assumption we've been making is that docked clocks will be around double the portable clocks, and going by the portable clock estimate in the video of 540MHz, we would expect docked to be around 1080MHz, so slightly faster than what was tested in the video. Also, the switch's lowest portable speed was 40% of the docked speed, so I decided to check that one too. Also I'm assuming DLSS will be rendered concurrently with the next frame.

Here are a few clock speed and DLSS frametime combinations I've calculated:

• DF video: 1000MHz clock, 18.3ms DLSS (4k30 viable, 4k60 impossible)
• 2x DF portable clock estimate: 1080MHz clock, 16.94ms DLSS (4k60 just barely impossible)
• Minimum clock speed for 16.7ms DLSS: 1095.8MHz
• Docked clock estimate from earlier in this thread: 1100MHz, 16.63ms DLSS (4k60 just barely possible)
• Docked clock 2.5x estimated portable clock (like switch): 1350MHz, 13.56ms DLSS (plenty of room for 4k60)

In conclusion; DLSS to 4k60 is impossible if the docked clock is 1GHz like the DF video assumed, but it doesn't take much more docked speed to make it theoretically viable.

13,56 is not "plenty of room" for 4k60. It means the game got about 3 ms to render before dlss.

 

[Cuzizkool]

okay, the DF video has put some of us a bit on edge, sure. But what about first party games? All this talk about graphically intensive 3rd party games, but what about the next 3D Mario/Zelda/Kart? Could Nintendo realistically make a 4k60 or 4k30 game using the set up Rich did in the video?

 

[Paul_Subsonic]

one more example of denoising costs, this time from Qualcomm's UE5 talk on ray traced shadows. their lumen talk had nothing on denoising costs

Sweet ! I hadn't found this one.
Although we have to keep in mind shadows are like the easiest RT thing to denoise.
We can't apply this denoising time to anything else.

 

[Molybdenum]

13,56 is not "plenty of room" for 4k60. It means the game got about 3 ms to render before dlss.

This is assuming concurrency, so DLSS would be upscaling frame n while the rest of the GPU is rendering frame n+1.

 

[CeramicPigeon]

okay, the DF video has put some of us a bit on edge, sure. But what about first party games? All this talk about graphically intensive 3rd party games, but what about the next 3D Mario/Zelda/Kart? Could Nintendo realistically make a 4k60 or 4k30 game using the set up Rich did in the video?

No, it would be technologically impossible with the clock speed used.
It would have to be higher, which would be feasible if this chip wasn't on 8nm.

 

[JoshuaJSlone]

In that case, I think I would prefer a 720p screen again
Running below native res just sucks IMO

Gonna have to go back to like a 2DS top screen if you really want to be safe from subnative resolutions. But that would make 95% of games look much worse.

I'm sorry if this sounds stupid but it'll be 2024 when this thing comes out. It needs to be able to hit 4K. Maybe not consistently, as not even PS5/XSX can, but it needs to at least be the ceiling. Something that 2D games could easily hit, or maybe remasters of older titles from 6th or 7th gen.

If it could physically output 4K, even base Switch could manage it in similarly edge cases.

 

[ItWasMeantToBe19]

I downloaded DS and realize almost immediately that I can't reach 99% GPU at even 240 FPS because I have a 3080 and this is very embarrassing so I'll try something else later if no one else does, lol.

 

[JoshuaJSlone]

okay, the DF video has put some of us a bit on edge, sure. But what about first party games? All this talk about graphically intensive 3rd party games, but what about the next 3D Mario/Zelda/Kart? Could Nintendo realistically make a 4k60 or 4k30 game using the set up Rich did in the video?

BOTW at 4K60 is often talked about as a big feat because it would be a ~dozen times as many pixels as the Switch version. But not all Switch games would be such a stretch. Take something like MK8 or Smash that's already 1080p60 (the vast majority of the time) and a 4K60 version without DLSS or any special tricks should be no problem. Building from such a base wouldn't leave a lot of spare oomph to improve things further, though.

 

[TomNookYankees]

A Plague Tale: Requiem: A game that didn't come to last generation consoles, that runs at 900p30fps on the Series S, is here comfortably at 1080p30fps.

But parts of Requiem had to be run at 900p to achieve stability in Rich's video though.

 

[Deleted member 887]

I'm sorry if this sounds stupid but it'll be 2024 when this thing comes out. It needs to be able to hit 4K. Maybe not consistently, as not even PS5/XSX can, but it needs to at least be the ceiling. Something that 2D games could easily hit, or maybe remasters of older titles from 6th or 7th gen.

Ahhh! Let me clarify.

When Rich said “4K is impossible” he was talking in the context of 4K DLSS, and specifically in the area of 60fps. In other words, would modern games get 4k60?

There is enough power in the device to skip DLSS and get older gen games to 4K. There is enough power to get Tears of the Kingdom to 4k, as long as you don’t mind staying on a 30fps update.

 

[Dekuman]

A (long) Note About The DF Video and How To Interpret The Results
I want to quote a DF Direct real quick (edited for clarity)

Rich: Bobby Kotick is saying that basically we're looking at last gen level performance, you know PS4 style. Which I think would be you know not bad for a mobile device but the point is I think that is only a fraction of what these new machines actually offerwouldn't you agree Oliver?

Oliver: Yeah I think it's an interesting quote that nonetheless doesn't totally align with what we think we know about the Switch 2 too because, just as one example, could a PS4 play a game like breath of the wild at an effective 4K 60 Resolution right which is the the rumor?[ ...] So in terms of the raw Hardware horsepower I would suspect the switch 2 could be falling in line with the capabilities of a PlayStation 4 or an Xbox one but uh I don't really think in terms of the actual rendering power the switch 2 might might bring to the table in terms of offering dlss uh bringing that to the table I think the switch 2 should should be able to outperform those consoles at least in scenarios where the technology is being effectively used

Rich touched on it again in his video yesterday at 8:18

 

[Gerald]

So If 660MHz in Handheald mode, 1.1GHz in Docked Mode, what think guys will be CPU clocks? 2GHz would be perfect imo

 

[Gerald]

A (long) Note About The DF Video and How To Interpret The Results
I want to quote a DF Direct real quick (edited for clarity)



Consoles are weird machines. CPUs and GPUs both have new major release every two years, and the industry is constantly pushing those boundaries. A console needs to provide a great bang for the buck now, while making decisions that will allow the technology to hold up 7 years from now.

This is the way to think about Rich's tests. Nvidia brings technologies to the table that have, even in their budget products, never been tested in such a small device. It brings up a number of questions, like -

• How good does a console implementation of DLSS - lower than 4k res, on a big TV - look?
• How well does DLSS perform on a low power device?
• How good is Nvidia's Ray Tracing hardware on a lower power device?
• What could such a small GPU do, if paired with a more modern CPU/Storage solution?

Let's look at the games and see what lessons we can extract instead of "this looks good" or "that runs poorly" or "I don't like that resolution."

Death Stranding: PS4 delivered 1080p30 on this game, and Rich shows that native. PS4 Pro delivered 4k checkerboarded, also 30fps, the test machine does it with DLSS Ultra Performance mode, though somewhat unstably. Rich also shows a 1440p30 mode that is much more stable.

What did we learn: The drum I have beat is "PS4 power, PS4 Pro experiences possible, but different tech means devs might make different decisions." Here we see all of that hold up. The raw power is plenty good enough to just "do" the PS4 experience without any real work. A 4k DLSS experience hits some performance snags that would be ironed out by a quality optimized port.

We also see the DLSS is a totally different technology from checkerboarding. It looks better, it's more flexible, but its cost grows differently. This creates different tradeoffs, and we shouldn't expect devs to make exactly the same choices.

Cyberpunk 2077: Death Stranding is a last gen console game with a good PC port. Cyberpunk is a PC game with a shitty last gen console port. Rich shows us the game running at PS5 quality settings, but at 1080p30, with instability.

What did we learn: We start to see how and why developers might make different decisions than on the AMD consoles. PS4 Pro runs at 1080p30, with a series of settings that are described by DF themselves as "extremely blurry and just kind of visually kind of glitchy". Series S has a 1080p60 mode. Both are 4 TFLOP machines. Here we see the 3 TFLOP Ampere card absolutely smack the pant off the PS4 Pro, running at comparable frame rates and resolutions, with substantially higher quality settings, but unable to deliver the Series S 1080p60, even with DLSS enabled.

DLSS is a different tech, it doesn't behave like "just more flops."

A Plague Tale: Requiem: A game that didn't come to last generation consoles, that runs at 900p30fps on the Series S, is here comfortably at 1080p30fps.

What did we learn: That the GPU isn't all that matters. Plague Tale is rough on the GPU, sure, but it's famously CPU limited. Pairing even this weak GPU with a modern CPU and storage solution, and suddenly 9th gen exclusives become viable.

Control: This runs at 30fps on the last gen machines, and kinda badly at that. 1080p on the PS4 Pro. Here it runs more stably, but same resolution, matching the PS5's settings.

What did we learn: Here, once again, we have the PS4 Pro performance/resolution experience via DLSS, but that's not the interesting story. What's interesting is that we're getting that level of performance with ray tracing enabled. Compare to the PS5 - these settings are matched, with PS5 at 1440p30fps. The Series S can't deliver an RT mode at all. But here we actually have a case where something "comparable" to the PS5's RT mode is actually easier to achieve than an ugly-but-fast 60fps mode.

Again, just like DLSS, RT cores change the math, opening up options that aren't possible on other hardware. RT is viable.

Fortnite: Rich tests with Nanite on, Lumen with hardware RT, Virtual shadow maps at high. With DLSS, he gets a comfortable 1080p30.

What did we learn: "Do I wanna play Fornite at 30fps???" I dunno man, I don't know your life. Who cares, that's not what this is about. The next-gen game engine, running with it's entire feature set fully enabled, and is still delivering HD resolutions and acceptable frame rates.

This is the power of a modern device. At nearly every level, Nvidia is providing a more advanced solution than AMD. UE5 is built for temporal upscaling, and DLSS is best-of-breed. Nanite uses mesh shaders on hardware that supports it, and Ampere does. Lumen has a software fallback for better performance and older machines, but Nvidia's RT hardware performs nearly identically to the software solution.

"Is the Series S holding gaming back" is a dumb discussion point of the last few years. Now we get to ask "Is the PS5 holding gaming back?" With it's lack of mesh shaders, it's lack of decent hardware RT, it's lack of AI accelerated rendering - it's Nintendo that is making the promise of UE5 fully possible. And that's what Rich is demonstrating here.

Kinda offtopic but i think PS and Xbox will have their own upscaller based AI beacuse this is where is future and DLSS clearly show it, Maybe aleardy in PS5 Pro

 

[ILikeFeet]

Sweet ! I hadn't found this one.
Although we have to keep in mind shadows are like the easiest RT thing to denoise.
We can't apply this denoising time to anything else.

easy to denoise, but we see that it takes as much time as a high end device denoising more complex scenarios. on a battery powered device, there's gonna have to be some work in optimizing denoisers since they won't have the raw power. not that the work hasn't already been done judging by insider reports

it's a shame there's no Lumen detail because that's the one effect I have big questions for. the irradiance cache resolution and how Qualcomm mitigate a lot of the boiling effect in particular

 

[chocolate_supra]

A (long) Note About The DF Video and How To Interpret The Results
I want to quote a DF Direct real quick (edited for clarity)



Consoles are weird machines. CPUs and GPUs both have new major release every two years, and the industry is constantly pushing those boundaries. A console needs to provide a great bang for the buck now, while making decisions that will allow the technology to hold up 7 years from now.

This is the way to think about Rich's tests. Nvidia brings technologies to the table that have, even in their budget products, never been tested in such a small device. It brings up a number of questions, like -

• How good does a console implementation of DLSS - lower than 4k res, on a big TV - look?
• How well does DLSS perform on a low power device?
• How good is Nvidia's Ray Tracing hardware on a lower power device?
• What could such a small GPU do, if paired with a more modern CPU/Storage solution?

Let's look at the games and see what lessons we can extract instead of "this looks good" or "that runs poorly" or "I don't like that resolution."

Death Stranding: PS4 delivered 1080p30 on this game, and Rich shows that native. PS4 Pro delivered 4k checkerboarded, also 30fps, the test machine does it with DLSS Ultra Performance mode, though somewhat unstably. Rich also shows a 1440p30 mode that is much more stable.

What did we learn: The drum I have beat is "PS4 power, PS4 Pro experiences possible, but different tech means devs might make different decisions." Here we see all of that hold up. The raw power is plenty good enough to just "do" the PS4 experience without any real work. A 4k DLSS experience hits some performance snags that would be ironed out by a quality optimized port.

We also see the DLSS is a totally different technology from checkerboarding. It looks better, it's more flexible, but its cost grows differently. This creates different tradeoffs, and we shouldn't expect devs to make exactly the same choices.

Cyberpunk 2077: Death Stranding is a last gen console game with a good PC port. Cyberpunk is a PC game with a shitty last gen console port. Rich shows us the game running at PS5 quality settings, but at 1080p30, with instability.

What did we learn: We start to see how and why developers might make different decisions than on the AMD consoles. PS4 Pro runs at 1080p30, with a series of settings that are described by DF themselves as "extremely blurry and just kind of visually kind of glitchy". Series S has a 1080p60 mode. Both are 4 TFLOP machines. Here we see the 3 TFLOP Ampere card absolutely smack the pant off the PS4 Pro, running at comparable frame rates and resolutions, with substantially higher quality settings, but unable to deliver the Series S 1080p60, even with DLSS enabled.

DLSS is a different tech, it doesn't behave like "just more flops."

A Plague Tale: Requiem: A game that didn't come to last generation consoles, that runs at 900p30fps on the Series S, is here comfortably at 1080p30fps.

What did we learn: That the GPU isn't all that matters. Plague Tale is rough on the GPU, sure, but it's famously CPU limited. Pairing even this weak GPU with a modern CPU and storage solution, and suddenly 9th gen exclusives become viable.

Control: This runs at 30fps on the last gen machines, and kinda badly at that. 1080p on the PS4 Pro. Here it runs more stably, but same resolution, matching the PS5's settings.

What did we learn: Here, once again, we have the PS4 Pro performance/resolution experience via DLSS, but that's not the interesting story. What's interesting is that we're getting that level of performance with ray tracing enabled. Compare to the PS5 - these settings are matched, with PS5 at 1440p30fps. The Series S can't deliver an RT mode at all. But here we actually have a case where something "comparable" to the PS5's RT mode is actually easier to achieve than an ugly-but-fast 60fps mode.

Again, just like DLSS, RT cores change the math, opening up options that aren't possible on other hardware. RT is viable.

Fortnite: Rich tests with Nanite on, Lumen with hardware RT, Virtual shadow maps at high. With DLSS, he gets a comfortable 1080p30.

What did we learn: "Do I wanna play Fornite at 30fps???" I dunno man, I don't know your life. Who cares, that's not what this is about. The next-gen game engine, running with it's entire feature set fully enabled, and is still delivering HD resolutions and acceptable frame rates.

This is the power of a modern device. At nearly every level, Nvidia is providing a more advanced solution than AMD. UE5 is built for temporal upscaling, and DLSS is best-of-breed. Nanite uses mesh shaders on hardware that supports it, and Ampere does. Lumen has a software fallback for better performance and older machines, but Nvidia's RT hardware performs nearly identically to the software solution.

"Is the Series S holding gaming back" is a dumb discussion point of the last few years. Now we get to ask "Is the PS5 holding gaming back?" With it's lack of mesh shaders, it's lack of decent hardware RT, it's lack of AI accelerated rendering - it's Nintendo that is making the promise of UE5 fully possible. And that's what Rich is demonstrating here.

 

[SiG]

Kinda offtopic but i think PS and Xbox will have their own upscaller based AI beacuse this is where is future and DLSS clearly show it, Maybe aleardy in PS5 Pro

I really doubt it, unless AMD pivots their entire GPU line to suddenly support heavy machine learning loads by having dedicated hardware for Tensor-like instructions that can run in parallel with shaders like Nvidia's (and even perhaps Intels) GPUs.

 

[ItWasMeantToBe19]

The big question is whether or not the Switch 2 has enough RAM for a TotK port to display HD textures.

 

[darthdiablo]

I really doubt it, unless AMD pivots their entire GPU line to suddenly support heavy machine learning loads by having dedicated hardware for Tensor-like instructions that can run in parallel with shaders like Nvidia's (and even perhaps Intels) GPUs.

Yup. In other words, non-nvidia (ie: AMD) GPUs would rely on software-based upscaling technology, while nvidia GPUs (at least RTX, not GTX line) rely on (clearly superior) hardware-based upscaling technology.

 

[Concernt]

The way Nvidia and mobile development goes, I would be surprised if the next generation was, maybe not short, but complemented with new hardware releases regularly, and a software continuum of sorts. Like say they go with numbers.

2 in 2024, 3 in 2027, etc.

Software says "Nintendo Switch 1 and up", "2 and up", etc. to indicate compatibility.

In a funny way, this is the directed they headed with DS, with DS, DSi, 3DS and New 3DS being a similar setup, only with very few, very minor enhancements by playing it on a more recent bit of hardware.

Kinda offtopic but i think PS and Xbox will have their own upscaller based AI beacuse this is where is future and DLSS clearly show it, Maybe aleardy in PS5 Pro

I believe we've seen in the leak from the court case that Microsoft's hardware plan is to grit their teeth and ride this generation out without a Pro version, but they're explicitly exploring AI upscaling and ARM for NG Xbox, so we could see Nvidia enter a new console arena IF they decided to go with both of those ideas, AND choose Nvidia.

 

[Tof]

No, it would be technologically impossible with the clock speed used.
It would have to be higher, which would be feasible if this chip wasn't on 8nm.

Most here don’t think it will be on 8nm despite what the DF video claims. I can’t find the post Thraktor wrote some time ago but it’s very comprehensive in breaking down why 8nm is unlikely to be the the node that NG uses. To summarise I believe it boils down to battery life considerations and reducing the voltage below a certain value actually causes instability and less optimal performance when on 8nm however if on 5nm the maths works out just fine. I probably massacred the explanation somewhere but I’m on mobile and can’t find the original explanation anywhere.

 

[Gerald]

I really doubt it, unless AMD pivots their entire GPU line to suddenly support heavy machine learning loads by having dedicated hardware for Tensor-like instructions that can run in parallel with shaders like Nvidia's (and even perhaps Intels) GPUs.

Consoles SoC are aleardy very custom( PS more than Xbox but still) and they add their own stuff to it, you can see from patents that Cerny work on PS own RT Solution( very likely aleardy for Pro) or that Cerny work on AI Based Upscaller is Also on patent, Xbox Also work on that what we can see on slide from FTC case court but their solution (at least According to slide) is to add NPU to SoC(chip responsible only for their own AI based Upscaller and Frame Gen) alongside CPU and GPU

 

[darthdiablo]

Most here don’t think it will be on 8nm despite what the DF video claims. I can’t find the post Thraktor wrote some time ago but it’s very comprehensive in breaking down why 8nm is unlikely to be the the node that NG uses. To summarise I believe it boils down to battery life considerations and reducing the voltage below a certain value actually causes instability and less optimal performance when on 8nm however if on 5nm the maths works out just fine. I probably massacred the explanation somewhere but I’m on mobile and can’t find the original explanation anywhere.

Might be the post by Thraktor you were thinking of? I keep it in my bookmarks

 

[Tof]

Might be the post by Thraktor you were thinking of? I keep it in my bookmarks

Much obliged!

 

[Gerald]

The way Nvidia and mobile development goes, I would be surprised if the next generation was, maybe not short, but complemented with new hardware releases regularly, and a software continuum of sorts. Like say they go with numbers.

2 in 2024, 3 in 2027, etc.

Software says "Nintendo Switch 1 and up", "2 and up", etc. to indicate compatibility.

In a funny way, this is the directed they headed with DS, with DS, DSi, 3DS and New 3DS being a similar setup, only with very few, very minor enhancements by playing it on a more recent bit of hardware.


I believe we've seen in the leak from the court case that Microsoft's hardware plan is to grit their teeth and ride this generation out without a Pro version, but they're explicitly exploring AI upscaling and ARM for NG Xbox, so we could see Nvidia enter a new console arena IF they decided to go with both of those ideas, AND choose Nvidia.

Also in slide was nothing about go with Nvidia, still AMD but a lot their custom work and add NPU to SoC(chip only responsible for AI Stuff)There was also leak from AMD reliable leaker Kepler_L2, that slide might be outdated and Xbox will release next gen in 2026, before PlayStation, we will can see then that next Xbox would be Zen5/RDNA5, and PS6 Zen6/RDNA6

 

[redmutineer75]

Anyone wanna do avatar bets on whether or not Nintendo will acknowledge the new system on the day before or of the investor meeting on the 7th? I'll be your bookie.