Nintendo Switch 2 will have the best upscaling technology of any consoles
- Thread starter Buggy Loop
- Start date
FIREKNIGHT2029
Member
yurinka
Member
Last edited:
tkscz
Member
One game after only being worked on for 7 weeks was at 540p in handheld mode. The rest were all at 720p to 1080p in handheld mode.Yes, PS4 games running at 540p 30fps will look great upscaled to 4K
DenchDeckard
Moderated wildly
I would not bet on this. I'm sure it will be good but it is just a small device. I don't think it will have too many tensor cores packed in.
Here is the thing!The announcement from last night, the move to change from old DLSS Convolutional Neural Networks (CNN) to transformer model is applicable for all RTX hardware
See ENHANCED: DLSS Super Resolution.
Even Switch 2 with rumoured old ass Ampere will have the best upscaling tech of console world
Next gen, if a console manufacturer goes with AMD, the other should go with Nvidia, the one with the deeper pockets. It's getting ridiculous how good Nvidia is getting at AI.
At Switch 2 level of graphics PS5 & Xbox Series X don't even need to be upscaled
Audiophile
Member
540>1080 DLSS Transformer Performance > Switch 2 Display | Portable
720>1440 DLSS Transformer Performance > Scale to 1080/2160p External Display | Docked
Might be reasonable..?
Plus the screen is 120 VRR, so with VRR at the upper end and LFC+VRR at the lower end, performance slips or even just sitting out of the traditional 30/40/60 brackets for prolonged times on the switch display itself won't really matter as much.
Either way, DLSS + VRR seem like absolute no brainers for a power-limited device that's primarily used with a fixed screen that comes as standard which they can define.
Even though DLSS incurs its own performance/power cost on top of the base ~25% of a 540p frame vs a 1080p frame, I'd still expect it to come in at around half the cost for a frame vs native 1080p.
720>1440 DLSS Transformer Performance > Scale to 1080/2160p External Display | Docked
Might be reasonable..?
Plus the screen is 120 VRR, so with VRR at the upper end and LFC+VRR at the lower end, performance slips or even just sitting out of the traditional 30/40/60 brackets for prolonged times on the switch display itself won't really matter as much.
Either way, DLSS + VRR seem like absolute no brainers for a power-limited device that's primarily used with a fixed screen that comes as standard which they can define.
Even though DLSS incurs its own performance/power cost on top of the base ~25% of a 540p frame vs a 1080p frame, I'd still expect it to come in at around half the cost for a frame vs native 1080p.
Last edited:
Kataploom
Gold Member
Even if you were right, which I don't think is gonna be the case, when OP made the thread there was no FSR 4 available or confirmed for PS5 so want lying, and even so right now Switch 2 will have the best upscalers for a while until PS5 Pro gets FSR4
RoadHazard
Gold Member
PSSR is better than the DLSS the Switch 2 can handle I'm pretty sure. You're not getting full blown DLSS4 on Switch, it just doesn't have the hardware for it.
CitizenKratos
Member
I just changed my vote to “OP is crazy”. I was originally bullish on Switch 2 DLSS but I think PS5 Pro will easily edge out the Switch 2 solution once PSSR gets its big (FSR4 type) update in 2026 if not at some point this year like is being foreshadowed by the Assassins Creed Shadows dev. For Switch 2 I think the TOPs won’t be high enough to compete with most PC Card DLSS solutions nor the high TOPs in PS5 Pro.
Kataploom
Gold Member
It's not even better than DLSS 2 lol
64gigabyteram
Reverse groomer.
yeah it'll have the best upscaling tech... It just doesnt really matter that much in the grand scheme of things. DLSS is there to clean up a mediocre TAA image, it doesnt suddenly give you graphics 2x better than what you can actually render.
Gambit2483
Member
Switch 2 DLSS will make PSSR look like amateurish 3rd rate ghetto upscaling.
bender
What time is it?
[Verse 1]
As the polygon draws
On a cold and gray Chicago mornin'
A poor up scaling solution is born
In the ghetto (In the ghetto)
And Mark Cerny cries
'Cause if there's one thing that he don't need
It's another hungry upscaler to feed
In the ghetto (In the ghetto)
[Bridge]
People, don't you understand
PSSR needs a helping hand
Or that learning model'll grow to be an angry upscaller someday
take a look at you and me
Are we too blind to see
Do we simply turn our heads
And look the other way?
[Verse 2]
Well, the world turns
And a hungry little model with a noisy glow
Plays on the screen as the cold wind blows
In the ghetto (In the ghetto)
And his hunger burns
So he starts to draw streets at night
And he learns how to steal and he learns NVidia's might
In the ghetto (In the ghetto)
[Bridge]
Then one night in desperation
The learaning model breaks away
He buys a gun, he steals a car
Tries to run, but he don't get far
[Verse 3]
And as Mark Cerny cries
As a crowd gathers 'round an angry learning model
Facedown on the street with a gun in his hand
In the ghetto (In the ghetto)
And as Mark Cerny's learning model dies (In the ghetto)
On a cold and gray Chicago mornin'
Another learning model is born
In the ghetto (In the ghetto)
[Outro]
And Mark Cerny cries
In the ghetto (In the ghetto)
(In the ghetto)
Topher
Identifies as young
sendit
Member
[Verse 1]
As the polygon draws
On a cold and gray Chicago mornin'
A poor up scaling solution is born
In the ghetto (In the ghetto)
And Mark Cerny cries
'Cause if there's one thing that he don't need
It's another hungry upscaler to feed
In the ghetto (In the ghetto)
[Bridge]
People, don't you understand
PSSR needs a helping hand
Or that learning model'll grow to be an angry upscaller someday
take a look at you and me
Are we too blind to see
Do we simply turn our heads
And look the other way?
[Verse 2]
Well, the world turns
And a hungry little model with a noisy glow
Plays on the screen as the cold wind blows
In the ghetto (In the ghetto)
And his hunger burns
So he starts to draw streets at night
And he learns how to steal and he learns NVidia's might
In the ghetto (In the ghetto)
[Bridge]
Then one night in desperation
The learaning model breaks away
He buys a gun, he steals a car
Tries to run, but he don't get far
[Verse 3]
And as Mark Cerny cries
As a crowd gathers 'round an angry learning model
Facedown on the street with a gun in his hand
In the ghetto (In the ghetto)
And as Mark Cerny's learning model dies (In the ghetto)
On a cold and gray Chicago mornin'
Another learning model is born
In the ghetto (In the ghetto)
[Outro]
And Mark Cerny cries
In the ghetto (In the ghetto)
(In the ghetto)
Someone needs to start a Switch 2 Power King Memes thread, similar to: https://www.neogaf.com/threads/ot-xbox-power-king-memes.1531914/
Something we can look back on and say "this was retarded"
Last edited:
bender
What time is it?
Gambit2483
Member
[Verse 1]
As the polygon draws
On a cold and gray Chicago mornin'
A poor up scaling solution is born
In the ghetto (In the ghetto)
And Mark Cerny cries
'Cause if there's one thing that he don't need
It's another hungry upscaler to feed
In the ghetto (In the ghetto)
[Bridge]
People, don't you understand
PSSR needs a helping hand
Or that learning model'll grow to be an angry upscaller someday
take a look at you and me
Are we too blind to see
Do we simply turn our heads
And look the other way?
[Verse 2]
Well, the world turns
And a hungry little model with a noisy glow
Plays on the screen as the cold wind blows
In the ghetto (In the ghetto)
And his hunger burns
So he starts to draw streets at night
And he learns how to steal and he learns NVidia's might
In the ghetto (In the ghetto)
[Bridge]
Then one night in desperation
The learaning model breaks away
He buys a gun, he steals a car
Tries to run, but he don't get far
[Verse 3]
And as Mark Cerny cries
As a crowd gathers 'round an angry learning model
Facedown on the street with a gun in his hand
In the ghetto (In the ghetto)
And as Mark Cerny's learning model dies (In the ghetto)
On a cold and gray Chicago mornin'
Another learning model is born
In the ghetto (In the ghetto)
[Outro]
And Mark Cerny cries
In the ghetto (In the ghetto)
(In the ghetto)
Buggy Loop
Gold Member
No reason DLSS 4 is not implemented. DLSS 4 has been tested on the lowest laptop ampere models and even downclocks and works. In fact the drivers that dropped after blackwell launch had transformer model catching up to old CNN model performances but the thing is CNN model also got a boost.
DLSS 4 allows performance to look better than even DLAA CNN sometimes and definitely quality, but with a performance boost. Absolutely no reason to not use it in dock if you want 4k or 1440p.
Why would it be upscaling from a lower resolution? That's game dependent. This thread is not to claim that Switch 2 runs PS5 games, so that comment doesn't make sense.
The upscaler res for res will beat FSR 4 as its already known. And imagine, it needs a pro console, base consoles are left with fucking nothing, FSR 1 & 2?
Nvidia sees a LOT of improvement on horizon for transformer model too. This will evolve.
Will I have to bump this again for the inevitable day DLSS 4 is used on switch? Because I will.
Woopah
Member
It has 48.
Imtjnotu
Member
And people think it will outperform thr pro in here....
SweetTooth
Gold Member
Buggy Loop
Gold Member
And I really hope that PSSR does not use all of PS5 Pro's TOPs, because otherwise the tech is a joke.
(and no, I don't think it uses all TOPs, this has been discussed many times)
DLSS 4 is running on Turing cards that are in under 100 TOPs range and of course its not using all tensor cores all at once. Never in a render graph you'll see 100% utilization for upcaling.
Funny thing you ask
Peoples so far are benching with Ampere equivalents, which is GA10x SM (RTX 30x0 series)
Jetson Orin while is "ampere" is not GA10x of course. Just like a GA100 SM for an A100 is not equivalent to GA10x gaming cards.
The GA10 version Orin uses is different than GA10x.
For AI, because Jetson Orin was mainly focus for AI automotive industry, it actually has the double sized tensor core config found on GA100 used for AI datacenters. Not what Ampere 30x0 series use for gaming.
So Orin has 256 FP16 FMA Ops like GA100. GA10b Orin is actually.
The AGX Orin, 64 tensor cores
1.3Ghz(64GB dev kit) * 64 tensor cores * 256 FP16 FMA * 2 (FMA to OPs) = 42.5 TFLOPs
INT8 computing capability should be 42.5 * 2 = 85 TOPs.
Sparse INT8 is double so 85 * 2 = 170 TOPS and this confirms the AGX Orin whitepaper. Good, one thing established.
So in docked mode where DLSS makes sense to upscale, @ the T239 leaked specs of 48 tensor cores, we're at ~100-125 TOPS depending on clock. (~1GHz leaked many times including DF)
So even taking the laptop 2050 (ampere and not turing if I have to refresh memories) which has shown to be having DLSS 4
We have 735 MHz base clock, 1245 MHz boost clock. Let's go optimistic with 1245 MHz for fun. 64 Tensor cores but GA10x tensor versions
1.245Ghz * 64 tensor cores * 128 FP16 FMA * 2 (FMA to OPs) = 20.4 TFLOPs
INT8 computing capability should be 20.4 * 2 = 40.8 TOPs.
Sparse INT8 is double so 85 * 2 = 81.6 TOPS
~80% of a docked Switch 2, and this is without counting on DLA which adds a ton of TOPs, but even without those
Richard from DF used 2050 @ 750MHz to try to match an equivalent SM T239, so he got 49.2 TOPs... oooopppppssssieeeee Richie
Now here is that ~80 TOPs laptop 2050 with DLSS 4
DF dipped from 720p native 32fps to 1080p DLSS quality 29 fps and made a whole case against Switch 2 using DLSS!
While with 80% of Switch 2 TOPS, the above video still shows that from 1080p native to 1080 DLSS quality it climbs in performance. WITH DLSS 4 WHICH IS MORE DEMANDING THAN RICH's CNN 2050 VIDEO! Switch 2 is at minimum 2 times TOPS of what DF did in that video.
Because peoples downclocking a 3050 or a 2050 to match leaked specs were not accounting for the GA10b features of Orin and DF is guilty of this. They not only had the stupid idea of taking a GPU that is memory matched for a higher number of SMs than T239, thus when going under clocks to try to match the SM difference they choked it under 25GB/s TFlops that Ampere likes, but also 4GB VRAM starved and they choked it in TOPs because they forgot these kind of very important details.
With Nvidia themselves announcing DLSS on Switch 2, its a no brainer.
Last edited:
OmegaSupreme
advanced basic bitch
It has dlss so yes. Fsr is dog shit on consoles and only just got decent on PC with fsr4 which isn't on consoles.
Kataploom
Gold Member
And I really hope that PSSR does not use all of PS5 Pro's TOPs, because otherwise the tech is a joke.
(and no, I don't think it uses all TOPs, this has been discussed many times)
DLSS 4 is running on Turing cards that are in under 100 TOPs range and of course its not using all tensor cores all at once. Never in a render graph you'll see 100% utilization for upcaling.
Funny thing you ask
Peoples so far are benching with Ampere equivalents, which is GA10x SM (RTX 30x0 series)
Jetson Orin while is "ampere" is not GA10x of course. Just like a GA100 SM for an A100 is not equivalent to GA10x gaming cards.
The GA10 version Orin uses is different than GA10x.
For AI, because Jetson Orin was mainly focus for AI automotive industry, it actually has the double sized tensor core config found on GA100 used for AI datacenters. Not what Ampere 30x0 series use for gaming.
So Orin has 256 FP16 FMA Ops like GA100. GA10b Orin is actually.
The AGX Orin, 64 tensor cores
1.3Ghz(64GB dev kit) * 64 tensor cores * 256 FP16 FMA * 2 (FMA to OPs) = 42.5 TFLOPs
INT8 computing capability should be 42.5 * 2 = 85 TOPs.
Sparse INT8 is double so 85 * 2 = 170 TOPS and this confirms the AGX Orin whitepaper. Good, one thing established.
So in docked mode where DLSS makes sense to upscale, @ the T239 leaked specs of 48 tensor cores, we're at ~100-125 TOPS depending on clock. (~1GHz leaked many times including DF)
So even taking the laptop 2050 (ampere and not turing if I have to refresh memories) which has shown to be having DLSS 4
We have 735 MHz base clock, 1245 MHz boost clock. Let's go optimistic with 1245 MHz for fun. 64 Tensor cores but GA10x tensor versions
1.245Ghz * 64 tensor cores * 128 FP16 FMA * 2 (FMA to OPs) = 20.4 TFLOPs
INT8 computing capability should be 20.4 * 2 = 40.8 TOPs.
Sparse INT8 is double so 85 * 2 = 81.6 TOPS
~80% of a docked Switch 2, and this is without counting on DLA which adds a ton of TOPs, but even without those
Richard from DF used 2050 @ 750MHz to try to match an equivalent SM T239, so he got 49.2 TOPs... oooopppppssssieeeee Richie
Now here is that ~80 TOPs laptop 2050 with DLSS 4
DF dipped from 720p native 32fps to 1080p DLSS quality 29 fps and made a whole case against Switch 2 using DLSS!
While with 80% of Switch 2 TOPS, the above video still shows that from 1080p native to 1080 DLSS quality it climbs in performance. WITH DLSS 4 WHICH IS MORE DEMANDING THAN RICH's CNN 2050 VIDEO! Switch 2 is at minimum 2 times TOPS of what DF did in that video.
Because peoples downclocking a 3050 or a 2050 to match leaked specs were not accounting for the GA10b features of Orin and DF is guilty of this. They not only had the stupid idea of taking a GPU that is memory matched for a higher number of SMs than T239, thus when going under clocks to try to match the SM difference they choked it under 25GB/s TFlops that Ampere likes, but also 4GB VRAM starved and they choked it in TOPs because they forgot these kind of very important details.
With Nvidia themselves announcing DLSS on Switch 2, its a no brainer.
It's amazing, I work on software dev and all this just makes me feel junior lol, thanks for the explanation!
SweetTooth
Gold Member
And I really hope that PSSR does not use all of PS5 Pro's TOPs, because otherwise the tech is a joke.
(and no, I don't think it uses all TOPs, this has been discussed many times)
DLSS 4 is running on Turing cards that are in under 100 TOPs range and of course its not using all tensor cores all at once. Never in a render graph you'll see 100% utilization for upcaling.
Funny thing you ask
Peoples so far are benching with Ampere equivalents, which is GA10x SM (RTX 30x0 series)
Jetson Orin while is "ampere" is not GA10x of course. Just like a GA100 SM for an A100 is not equivalent to GA10x gaming cards.
The GA10 version Orin uses is different than GA10x.
For AI, because Jetson Orin was mainly focus for AI automotive industry, it actually has the double sized tensor core config found on GA100 used for AI datacenters. Not what Ampere 30x0 series use for gaming.
So Orin has 256 FP16 FMA Ops like GA100. GA10b Orin is actually.
The AGX Orin, 64 tensor cores
1.3Ghz(64GB dev kit) * 64 tensor cores * 256 FP16 FMA * 2 (FMA to OPs) = 42.5 TFLOPs
INT8 computing capability should be 42.5 * 2 = 85 TOPs.
Sparse INT8 is double so 85 * 2 = 170 TOPS and this confirms the AGX Orin whitepaper. Good, one thing established.
So in docked mode where DLSS makes sense to upscale, @ the T239 leaked specs of 48 tensor cores, we're at ~100-125 TOPS depending on clock. (~1GHz leaked many times including DF)
So even taking the laptop 2050 (ampere and not turing if I have to refresh memories) which has shown to be having DLSS 4
We have 735 MHz base clock, 1245 MHz boost clock. Let's go optimistic with 1245 MHz for fun. 64 Tensor cores but GA10x tensor versions
1.245Ghz * 64 tensor cores * 128 FP16 FMA * 2 (FMA to OPs) = 20.4 TFLOPs
INT8 computing capability should be 20.4 * 2 = 40.8 TOPs.
Sparse INT8 is double so 85 * 2 = 81.6 TOPS
~80% of a docked Switch 2, and this is without counting on DLA which adds a ton of TOPs, but even without those
Richard from DF used 2050 @ 750MHz to try to match an equivalent SM T239, so he got 49.2 TOPs... oooopppppssssieeeee Richie
Now here is that ~80 TOPs laptop 2050 with DLSS 4
DF dipped from 720p native 32fps to 1080p DLSS quality 29 fps and made a whole case against Switch 2 using DLSS!
While with 80% of Switch 2 TOPS, the above video still shows that from 1080p native to 1080 DLSS quality it climbs in performance. WITH DLSS 4 WHICH IS MORE DEMANDING THAN RICH's CNN 2050 VIDEO! Switch 2 is at minimum 2 times TOPS of what DF did in that video.
Because peoples downclocking a 3050 or a 2050 to match leaked specs were not accounting for the GA10b features of Orin and DF is guilty of this. They not only had the stupid idea of taking a GPU that is memory matched for a higher number of SMs than T239, thus when going under clocks to try to match the SM difference they choked it under 25GB/s TFlops that Ampere likes, but also 4GB VRAM starved and they choked it in TOPs because they forgot these kind of very important details.
With Nvidia themselves announcing DLSS on Switch 2, its a no brainer.
I almost took you seriously until you said "PSSR tech is joke"
Then writing copy pasted wall of text doesn't make anything you write credible.
Hear me out:
Switch 2 isn't getting anything advanced or noteworthy. If it gets DLSS its gonna be hindered by weak unbalanced hardware, its weak by design plus Nintendo is historically cheap and clueless. Going to Nvidia or AMD or NASA or even to some outer alien technology won't change this fact.
Accept it and move on.
OmegaSupreme
advanced basic bitch
Switch 2 is confirmed to have dlss. You'd know that if you actually read the announcement straight from Nvidia lmao. Which is objectively better than any console level fsr that your PS5 or Xbox has available to it.
Accept it and move on.
Buggy Loop
Gold Member
If PSSR would take all the PS5 Pro TOPS, then it would be a joke
Then I said further down
(and no, I don't think it uses all TOPs, this has been discussed many times)
Because of fucking course PSSR did not eat up all the TOPs that Cerny presented. I'm responding to the dude saying that bigger TOPs make all the difference in upscaling. Not so much, as proven by the wide range of RTX TOPs performance running the upscaling.
PS5 Pro has leeway for other neural features. Peoples quoting the TOPs figure for PSSR are not very tech literate.
SweetTooth
Gold Member
Which version, I bet its the BAD one.
OmegaSupreme
advanced basic bitch
1080p? 1440p? That's what df has seen so far.
Gamer79
Predicts the worst decade for Sony starting 2022
If trolling you are doing a great job.. If serious there is no way. Did you see how bad Street figher 6 looked on Switch 2, it was a jaggy mess. Reports are that even cyberpunk on Switch looks soft and may not be that impressive.
SweetTooth
Gold Member
Its cute how Ninty kids stared using the words 4K, upscale, DLSS and such
Your mind will be blown once you start talking about RT in 2046
OmegaSupreme
advanced basic bitch
Haven't you embarrassed yourself enough in this thread?
SweetTooth
Gold Member
LordOcidax
Member
Reports are saying otherwise…According to some people the SF6 on Switch 2 are better than the Series S textures and the Cyberpunk build is just 7 weeks old…
Bernardougf
Member
ResilientBanana
Member
Nothing looks great when upscaled beyond comprehension, especially upscaled from a modest, low powered device.
What we’ve seen thus far has looked good, but in 5 years, the only thing that will look good are Nintendo Exclusives. But even then, expect frame dips like what we’ve experienced already halfway through the original Switch’s existence.
What we’ve seen thus far has looked good, but in 5 years, the only thing that will look good are Nintendo Exclusives. But even then, expect frame dips like what we’ve experienced already halfway through the original Switch’s existence.
Parazels
Member
We saw lots of native 4k games on PS5 and Xbox at launch as well. But then the base resolution significantly dropped.
Robb
Gold Member
Yeah, I can’t imagine 4k/120fps being relevant for anything more than the early cross-gen titles/upgrades and some smaller/indie games.
We’ll be back at sub 720p/1080p in no time.
ArtHands
Thinks buying more servers can fix a bad patch
My Steam library has more games that support RT than the entire RT supporting games on Playstation 5.
SweetTooth
Gold Member
I had to check the thread title which has no mention of either Steam or Playstation..so Im getting this vibe from your post:
Horseganador
Member
We'll see how he uses it, because so far it hasn't been seen in any game...
It's not confirmed what kind of Tensor cores will be in the T239. We can't assume everything will carry over, because Nintendo can decide to cut features to save die space. Also for the laptop test we're talking about a 6.6% difference in memory bandwidth (96 GB/s vs. 102.4 GB/s) at the leaked clocks.
Buggy Loop
Gold Member
But 2050 has more SMs and the downclock also impacted memory bandwidth. They do not respect the usual 25GB/s TFlops of Ampere.
For the tensor cores, at that point, why even pick T239 based on T234 according to rumors if its to change the fundamentals of what the SM is for that architecture. If we assume we cannot take what AGX Orin T234 is as a baseline, then anything goes for speculations on what this chipset is capable of.
I don't think DF touched the memory clocks, so it should still have 96 GB/s for 3.1 TF. I guess the lowered clock speed would affect the speed of the internal caches, so that is a factor.
Because Nvidia doesn't have a consumer Ampere based SoC? If you're replacing the CPU, removing the DLA, adding a media encoder from Ada as well as the improved clock gating and adding a file transfer engine, then I don't see why the Tensor cores are somehow "sacred". The PS5 APU was "based" on RDNA 2, but seems to ship with the front end from RDNA 1. While the PS5 Pro APU uses the CUs from RDNA 2, the geometry engine from RDNA 4, the RT units from RDNA 4 and custom hardware for upscaling. So no, I don't think saying that the T239 was based on the T234 tells us anything about the capabilities of the hardware, since Nintendo are free to request what they want.
Last edited: