OP
OP
Consumer Blackwell GPUs (Ada Lovelace-Next) are not expected until 1H 2025 at the earliest. Only datacentre Blackwell GPUs (Hopper Next) are expected to be released in 2024.
-
Hey everyone, staff have documented a list of banned content and subject matter that we feel are not consistent with site values, and don't make sense to host discussion of on Famiboards. This list (and the relevant reasoning per item) is viewable here.
StarTopic Future Nintendo Hardware & Technology Speculation & Discussion |ST| (Read the staff posts before commenting!)
- Thread starter Dakhil
- Start date
Kise Ryota
Bob-omb
I don't mind going against the expectations from this thread tbh lol
If we end up getting higher clocks while also having decent battery life, great! But I prefer to just expect the worst case scenario (IMO).
The last time I was hyped for the Switch's SoC using TSMC 16nm. This time I'll do it differently lol
Even though the GPU tests from the Nvidia hack were just tests, I would imagine they are still significant as to what Nvidia/Nintendo were going for. As it is, clocks we saw presented were 660Mhz (4.2W), 1.125Ghz (9.3W) and 1.38Ghz (12.0W). The question is, could the T239 fabricated on 8N even make that possible? Been looking over the existing supported modes and power efficiency section of the Nvidia docs related to the varying Jetson Orin series, and honestly, I simply don't see it.
ItWasMeantToBe19
Manakete
Is there any info about neural network based global illumination and whether that could be viable on the Switch 2?
Z0m3le
Bob-omb
That question is academic in nature. It just doesn't have a real barring on anyone here, the only people this matters to is the people fabbing the chip.Even though the GPU tests from the Nvidia hack were just tests, I would imagine they are still significant as to what Nvidia/Nintendo were going for. As it is, clocks we saw presented were 660Mhz (4.2W), 1.125Ghz (9.3W) and 1.38Ghz (12.0W). The question is, could the T239 fabricated on 8N even make that possible? Been looking over the existing supported modes and power efficiency section of the Nvidia docs related to the varying Jetson Orin series, and honestly, I simply don't see it.
- Pronouns
- he / him
If we exclude that, we'd basically be back to "we can probably assume it's something from NVIDIA again", anything goes.
GrandDemand
Cappy
- Pronouns
- He/him
Are you referring to Neural Radiance Caching? https://research.nvidia.com/publication/2021-06_real-time-neural-radiance-caching-path-tracing
Skimming through the paper, they used a 3090 + i9-9900K to achieve the results they did. The technique seems very compute intensive. I'd say it's not viable on the Switch 2
ItWasMeantToBe19
Manakete
No, I’m talking about a neural network that receives scene data and outputs indirect lighting based on the data given. Basically just guessing where the rays would go without sending out rays that bounce.
TLZ
Like Like
Gotdatmoneyy
Boo
The logic suggests a chip that big probably on the Switch's power budget at 8nm is unlikely. But I still work from the assumption that they will just clock it down appropriately if it is 8nm.
But you still end up with a solid device if not spectacular. But it's just hard to see why you would make such a big chip to run it at Switch clocks.
I would think there would be some clock boosts over the original Switch. It's hard to see the CPU not floating in the 1.4 to 1.5 GHz range. It would be odd to get no clock boost on the CPU side even if it is small. It is hard to see the GPU not getting some uptick from general switch clocks. Or at least matching. That's 460 mHz in handheld mode and 768 MHz docked. But I would think docked would get an uptick in the 900s. 921 MHz was the max but never used docked profile of the original Switch and that seems right around right. So you get 1.4 Tflops and 2.8 Tflops
These seem like they should be doable on 8nm. You aren't really gonna lower than the 307 and 768 clocka on the GPU. There is just some level of base performance that has to exist.
That orin power consumption took makes it seem unlikely that even at Switch clocks the chip would be within Switch energy consumption at 8nm. That basically says 3 things. The tool is wrong and the power saving features they implemented gain them back the wiggle room at those clocks. That's good. The tool is right and it just can't be 8nm. That's also good. The tool is right and they just clock even lower. This would be bad but why even make a chip like this if it can't even math Switch clocks.
Basically, unless option 3 happens, we will get a nice piece of kit.
But you still end up with a solid device if not spectacular. But it's just hard to see why you would make such a big chip to run it at Switch clocks.
I would think there would be some clock boosts over the original Switch. It's hard to see the CPU not floating in the 1.4 to 1.5 GHz range. It would be odd to get no clock boost on the CPU side even if it is small. It is hard to see the GPU not getting some uptick from general switch clocks. Or at least matching. That's 460 mHz in handheld mode and 768 MHz docked. But I would think docked would get an uptick in the 900s. 921 MHz was the max but never used docked profile of the original Switch and that seems right around right. So you get 1.4 Tflops and 2.8 Tflops
These seem like they should be doable on 8nm. You aren't really gonna lower than the 307 and 768 clocka on the GPU. There is just some level of base performance that has to exist.
That orin power consumption took makes it seem unlikely that even at Switch clocks the chip would be within Switch energy consumption at 8nm. That basically says 3 things. The tool is wrong and the power saving features they implemented gain them back the wiggle room at those clocks. That's good. The tool is right and it just can't be 8nm. That's also good. The tool is right and they just clock even lower. This would be bad but why even make a chip like this if it can't even math Switch clocks.
Basically, unless option 3 happens, we will get a nice piece of kit.
Maybe it's just copium/acceptance or whatever, but I'm probably gonna stop worrying too much about the possibility of S2 being 8nm. On paper, the Wii U had a gpu that was higher clocked and two gens ahead of the 360's gpu, with 22 MB more eDRAM and double the video RAM. Yet, while the Wii U had quite a few fantastic looking games, not one single game on the Wii U, came off as noticeably far off from what the 360 (and PS3) seemed to be capable of.
Similarly, the power difference between the two nodes for the S2 might not be that large when it comes to the final result when all is said and done.
Similarly, the power difference between the two nodes for the S2 might not be that large when it comes to the final result when all is said and done.
omniryu
Like Like
- Pronouns
- He/him
What was the name of this topic again?
Crazy it seems that speculation is almost on topic.
I mean what copium is there? That it has a more powerful CPU than the PS4 and PS4 Pro? It has an more advanced GPU and it can do DLSS. A Rendering technique to get you several fold better than the native render? Even if all the research of what nm the SOC is on happen to be a fluke. We are getting a more powerul machine than last gen consoles.
Z0m3le
Bob-omb
No this was just me miscommunicating, I thought he had estimated the GPU clock based on our expectations for docked power consumption. Such as 8w-10w for the docked GPU.
That's a good point. If it ends up being the clocks we want, everyone here will think 4nm is confirmed, while MLID will say "See, it's exactly what I told you. I was right about 8nm".
i think it will be quite competitive compared to other portable tech for lot longer time than Switch was.
unlike Switch it will (probably) launch at a very good node, and tech advancements are slowing down. It's almost impossible that we get efficiency gains comparable to Mariko during the Switch 2 gen.
Zedark
The guy with the ToV avatar
Short version: never underestimate the software feature set being able to future proof older hardware.
Longer version: You probably remember the impossible Switch ports. Games like Witcher 3 were eventually able to run on the Switch 1, but it took significant downgrades and significant frame rate issues in the biggest regions. But then, more recently, we have seen very impressive releases such as Nier Automata and Dying Light that showcase a much better visual comparison with PS4/XB1, without sacrificing much in terms of frame rates (Nier went from very unstable 40-60 fps on PS4/XB1 to a mostly stable 30 fps on Switch, and Dying light is 30 fps on all systems which is mostly stable according to DF). What happened to make the conversions so much more competent? The answer, in both cases, is TAAU: a temporal anti-aliasing method used to upscale the native resolution from a lower resolution to a higher one. It allows Switch to run at a lower internal resolution, which eases the rendering cost, and to let the upscaling method produce a competently upscaled image with much less rendering cost. This is how Switch 1 got so close to PS4/XB1 in these games. Now, TAAU is a pretty new method, and was not available for older ports. So what you see here is that software advances allowed for Switch 1 to punch above its previous weight.
Switch 2 is believed from the get-go to be able to punch above its weight. DLSS is the main feature discusses for this, but also in terms of ray-tracing it has an advantage of using dedicated RT cores and using ray reconstruction (a machine learning-based method for improving ray tracing quality). Some her even believe that Switch 2 could have RT capabilities on par with PS5. More importantly, though, for gen 9 games, the Switch 2 has the capacity to run at lower native resolutions than PS5/XSX and use DLSS to get a very competent output image, kind of similar to how TAAU narrowed the raw performance gap by smart use of upscaling software. Of course, PS5 and XSX also use FSR for upscaling, but DLSS has a better quality output and especially works better at lower native resolutions. And of course expectations for Switch 2 aren't that it matches PS5/XSX resolutions anyway. So for gen 9, Switch 2 is in a very good position.
Now the question is: what about gen 10 when it comes around in late 2028? Here, we have to guess a lot. Firstly, we don't know what raw specs are reasonable, since node shrinks are providing less and less improvement, and it might be that the raw performance gap will be smaller than the gen 8 -> gen 9 transition. However, at the same time we should probably expect AMD to have dedicated AI hardware in their systems by then, and either AMD or Playstation and Microsoft themselves will develop a more competent AI upscaler similar to DLSS, which is better than FSR. So it is unclear how much of an advantage there will be left based on the AI algorithms like DLSS and ray reconstruction that Switch 2 will have. The same question mark is there for ray tracing: how much worse will Switch 2 be at it compared to gen 10? Despite that, during 2028-2031, I would expect games to run on PS5/XSX/XSS still, and especially the latter system makes targeting Switch 2 in addition a bit easier. For sure there will be plenty of gen 10 games that won't run on Switch 2, and 2031-2033 will be worse in that regard.
So yeah, for gen 9 Switch 2 is very well positioned, but for gen 10 we simply don't know: are games going to target 8K, for example? Probably not, but if they are aiming for higher than 4K resolutions, then that gives Switch 2 a leg up in terms of getting gen 10 games (at much lower resolutions). In general, it is too early for us to say anything very definitive for gen 10, other than that it will be much more difficult to port such games than it would be for gen 9 games, of course.
Kise Ryota
Bob-omb
Let's say Drake is using TSMC 4N. That alone is by far the biggest improvement between Ampere and Ada. Backported from Ada we also have the clock gate thing (I don't know exactly what it is, but I know its benefit) and the media block (AV1). Only Ada's OFA would be missing. What other features Ada has that we probably wouldn't see making it into Drake's "Ampere" GPU?
I mean, at this point, why wouldn't they just use an Ada GPU?? That's something I never understood but never asked for opinions here. Are there any other features from Ada that would make it more expensive, and so they just opted for an "Ampere" GPU that's almost an Ada GPU?
I believe that's why the SEC 8nm never left my thoughts. Because, for me, knowing that Drake has an Ampere GPU (could someone remind me of where did we have this confirmation?) was always making that connection with the node. If they chose to just go with TSMC 4N, then I don't understand why not go with Lovelace anyway. Would Nvidia charge more just because it was their next architecture? And is there still a possibility of it being Lovelace somehow?
I mean, at this point, why wouldn't they just use an Ada GPU?? That's something I never understood but never asked for opinions here. Are there any other features from Ada that would make it more expensive, and so they just opted for an "Ampere" GPU that's almost an Ada GPU?
I believe that's why the SEC 8nm never left my thoughts. Because, for me, knowing that Drake has an Ampere GPU (could someone remind me of where did we have this confirmation?) was always making that connection with the node. If they chose to just go with TSMC 4N, then I don't understand why not go with Lovelace anyway. Would Nvidia charge more just because it was their next architecture? And is there still a possibility of it being Lovelace somehow?
ziggyrivers
thunder
- Pronouns
- He/him/his
Question: on whatever node does this end up on, will the Switch 2 be:
- Undocked: 1080p
- Docked: 4k
Steve
Koopa
- Pronouns
- he/him
Undocked we're thinking it'll br 1080 on handheld
Meanhwile 4K AI upscaled of dock, which would be great.
Overall i can see graphically demanding games will be using AI upscaled, like on
handheld 540p-1080p DLSS
Docked 1080p-4K DLSS
and we still don't know how native 4k will look like, maybe games like mario wonder will support native 4k.
I would say somewhere along the lines of:
- Undocked: 1080p DLSS quality
- Docked: 4k DLSS performance
NineTailSage
Bob-omb
It could be just the development time and R&D going into the NVN2 api.
Ampere and Lovelace really are so similar that Nvidia and Nintendo may have saw the differences as negligible. (Hence the backporting of Lovelace features)
I mean just look at PS5 vs Series X, both systems launched at the same time, the PS5 being some kind of hybrid of RDNA 1.5 with RDNA 2 features and the Series X having more native RDNA2 features. The PS5 at launch had more advanced developed devkits and tools while Microsoft struggled for awhile to get software running better on the platform.
So in theory them going with Ampere might have been to give developers plenty of time to get up to speed on the hardware, since even though T239 was taped-out in 2022. Ampere hardware has been around since 2020 and developers could just emulate a similar performance environment to T239 projected specifications and update their game engines accordingly...
Ampere and Lovelace really are so similar that Nvidia and Nintendo may have saw the differences as negligible. (Hence the backporting of Lovelace features)
I mean just look at PS5 vs Series X, both systems launched at the same time, the PS5 being some kind of hybrid of RDNA 1.5 with RDNA 2 features and the Series X having more native RDNA2 features. The PS5 at launch had more advanced developed devkits and tools while Microsoft struggled for awhile to get software running better on the platform.
So in theory them going with Ampere might have been to give developers plenty of time to get up to speed on the hardware, since even though T239 was taped-out in 2022. Ampere hardware has been around since 2020 and developers could just emulate a similar performance environment to T239 projected specifications and update their game engines accordingly...
There are some improvements to the RT and Tensor cores and cache structure, but by and large a 4nm ampere gpu with better power gating, should perform like a Lovelace gpu.
I imagine it had to do with timing, cost and practicality reasons. Drake is mostly designed in parallel with Orin by the same team. I imagine the 2 socs sharing architecture allowed them to design both of them quicker and cheaper.
HankyPanky
Like Like
- Pronouns
- Funcle
i know bad source and all that but the Moore is Dead guy said relevant things along these lines - a smaller node Lovelace option was offered but Nintendo went Orin 8nm because good enough/cheaper.
- Pronouns
- He/Him
Correct me if i'm wrong, but wouldn't nVidia be the one to decide on the line, not Nintendo?
HankyPanky
Like Like
- Pronouns
- Funcle
he said Nintendo were pitched both options by Nvidia and went with the cheapest one. now I know there is a lot of the smaller node is cheaper talk but we just don't know.
- Pronouns
- He/Him
From today, til Saturday.
Earliest i would feel people like Nate being ok to share something in case they heard something would be next two weeks.
Darknut85
#TeamRemake
I‘m ready for another delay./s
Kise Ryota
Bob-omb
If the differences are so negligible, it's weird to think they would have all this trouble of backporting features and porting an architecture to another node (from a different foundry) rather than just go with the Lovelace architecture in the first place. But thanks for giving your view on this.
Would it still be quicker and cheaper even though we are talking about SoC's being made on completely different nodes/foundries?
My guess is yes it would. But I have never designed a chip.
And why do you care, when performance wise the difference is probably negligible.
HankyPanky
Like Like
- Pronouns
- Funcle
technical question but i find it amusing we got dimensions for something directly related to the chip (substrate thing) and it doesn't rule out 8nm. for reference does anyone know the dimensions for this part in the OG Switch relative to the SOC?
NineTailSage
Bob-omb
Who said they were having trouble though?
All indications about T239 says it was taped-out around the same time as Lovelace, for all we know this might have been the plan all along...
So even though Orin and T239 started together, as Orin was finished T239 was still being worked on. This GPU jump is a massive one for Nintendo and game development is only getting longer and more complex, so any additional dev time avaliable is paramount.
Again I point to what Sony did with PS5, they could have easily waited for RDNA 2 but felt that porting some features and creating their own custom hardware would be more beneficial solution. If the Switch 2 delay rumors are even halfway true, then it sounds like enough software being ready was the main concern.
Last edited:
eye0h
"aggressively persue a multiplatform strategy"
- Pronouns
- He/Him
At this point it'd be really fucking funny
Eclissi
Moblin
Closed doors presentation part 2, revenge!
darthdiablo
+5 Death Stare
- Pronouns
- he/him
The problem also is MLID is assuming 8nm because other Orin GPUs are 8nm which is not really a safe assumption.
Tegra X1 (T210), a SoC with Maxwell-based GPU, was 20nm. The other Maxwell GPUs were 28nm.
It’s like saying it’s a safe assumption Tegra X1 will be 28nm only because the other Maxwell GPUs are 28nm. And we all know Tegra X1 ended up being 20nm.
Kise Ryota
Bob-omb
And that was exactly my point. If it's negligible, then why hold on Ampere (having to port it to a different foundry, which everyone here always says it's expensive) when you could just go with Lovelace? The biggest difference was always the node. I don't see a reason tbh, specially considering when the console will be (finally) released.
Porting Ampere's SEC 8nm to TSMC 4N is the trouble I'm referring to.
That's why I asked what other features could be so different between Ampere and Lovelace that would make sense (financially) to go with Ampere over Lovelace. It was always said here that having to design the chip on another node (and from a different foundry) was very expensive.
And about timing, devs could work on Ampere GPUs even if the final hardware would be Lovelace. It wouldn't make a difference for them. And for a console probably coming in 2025, even Lovelace wasn't too soon to work with.
ItWasMeantToBe19
Manakete
It will probably be
Undocked: 540p native upscaled to 1080p
Docked: 810p native upscaled to 1440p (then this image is upscaled to 4K by the TV)
For most titles
Kise Ryota
Bob-omb
Because Orin is also Ampere, does it make it easier to use the same architecture in terms of memory controller and other things that we'll see in a SoC but not in a PC gpu? Maybe that could make a difference on choosing the architecture.
We are outsiders looking in. Chip design is one of the most complex fields there is. I just assume there are thousands of little complexities that all adds up.
Memory controllers can be moved around between architectures. They're not bound by that but by node
Kise Ryota
Bob-omb
Then I really don't know lol
Doest make sense to me.
But whatever, I'll just accept we can't know the why and that's it. I don't want people here to get angry lol
Zedark
The guy with the ToV avatar
Yes, no, maybe, I don't know.
This is a per-game question, not a system question. Switch 2 will almost definitely support 1080p in handheld mode and 4K in docked mode. But whether many games will hit both of those targets consistently is still written in the stars. If you expect GTAVI to run at 4K, even after dlss, then you are probably off-base, but a new 3D Mario definitely might, or might not.
Point being, very few gen 9 games run at 4K on PS5/XSX, and Switch 2 will join that trend even for certain first party games imo. Third parties will hit below 4K regularly as well, and certain games might only hit 1080p after DLSS in docked mode (the more demanding ports). And all of that is to be expected, and will probably even look fine.
Melpomene
Ballistic
- Pronouns
- They/Them
I can see it now:
No new information from now until our presently expected release date.
The time arrives.
March, 2024...
And the reveal: someone leaks a new delay.
Ghostsonplanets
Ghost in the Sales
- Pronouns
- He/Him
See, that's the thing I think is unknown. Thraktor and Orin Power Tool resources are fantastic. But there's a lot of knobs that can be modified to achieve a better efficiency and we simply lack any info about it. Hence why I don't rule out 8N.
This is why I wanted to see a good testing of a RTX 2050 at really low clocks. RTX 2050 in the screenshot attached is drawing ~6 to 7W at 435MHz <-> 525MHz. GA10F compared to RTX 2050 lacks an entire second GPC, being much simpler and also has some Ada power savings. Wouldn't that mean that 8N is actually possible for T239 given the power draw x clocks showed in the screenshots? Possible? But I don't know, as we lack a good amount of data.
I don't know why this has come up once again, but the DLSS Test Tool clocks x power draw have 0 basis on reality and weren't tested at all on real hardware. They shouldn't be taken as an indicative of the clocks Nintendo will choose nor the power draw at such clocks.Even though the GPU tests from the Nvidia hack were just tests, I would imagine they are still significant as to what Nvidia/Nintendo were going for. As it is, clocks we saw presented were 660Mhz (4.2W), 1.125Ghz (9.3W) and 1.38Ghz (12.0W). The question is, could the T239 fabricated on 8N even make that possible? Been looking over the existing supported modes and power efficiency section of the Nvidia docs related to the varying Jetson Orin series, and honestly, I simply don't see it.
Lic already corrected Zom3le multiple times in the past about this matter. It's fine to believe and hype yourself, but it's not cool to drag others into a incorrect assumption.
darthdiablo
+5 Death Stare
- Pronouns
- he/him
ItWasMeantToBe19
Manakete
Resident Evil is on 9, not 10…
Last edited:
