StarTopic Future Nintendo Hardware & Technology Speculation & Discussion |ST| (Read the staff posts before commenting!)

[Deleted member 540]

BOTW2 is Still on Target for 2022.

I'll admit I was kind of hoping we find out it's getting pushed out to 2023 because it's going to be a launch title for the new system.

Why push it out to 2023? Well because it would be a bit shitty of Nintendo to launch new hardware in 2022 when the OLED had only been out for a year.

 

[Alovon11]

I'll admit I was kind of hoping we find out it's getting pushed out to 2023 because it's going to be a launch title for the new system.

Why push it out to 2023? Well because it would be a bit shitty of Nintendo to launch new hardware in 2022 when the OLED had only been out for a year.

Watch Nintendo Shadow drop the announcement of the Dane at the Game Awards for an early April Launch with BOTW2 getting its name there.

(At the very least we may get Metroid Prime 1 Remake/Remaster revealed at TGA if Emily Rodgers is to be believed on her "Finished Dev over the summer comment". There has to be some reason to hold it for that long, why not be a title alongside Dane)

Tinfoil Hat off

 

[BlackTangMaster]

Summary of the very few info we have on next switch hardware:

1) Ampere architecture
2) 8nm for node process
3) New SoC (not Tegra X1+)
4) Will have DLSS support via featuring Tensor Cores
5) (Very) Limited Ray Tracing support, few RT cores on SoC
6) Its still an hybrid console, power output limited to max 15W when docked
7) It will have a (few?) number of (third party only?) exclusive titles.

What we don’t know:
1) Console form factor, dimensions, weight, screen size, resolution and technology.
2) New joycons or will be the same
3) SoC specifics (GPU SMs, CPU core and which cores, clock, bus bandwidth)
4) RAM: amount, technology, speed, etc
5) Dock, the same as OLED model or different?

1) Ampere architecture : actually it could be Lovelace/Ampere/Turing or a custom version that will probably be named after the last revealed Desktop Architecture. If it is revealed in H2 2022, I would expect Nvidia to call it Lovelace even if it is 200% Pascal.

2) 8 nm : it hasn't taped out. N6 and 7LPP/5LPE/4LPX are still on the table.

5) I would not take Orin having RT cores as an evidence of T239 having RT cores. Nvidia won't need those cores on a shield TV.

6) They could target the Mariko 10W power budget especially if they are planning to release lite at the same time.

1) IMHO the console form factor was probably the OLED form factor before the chip shortages cutting Nintendo business case before signing the contract for the high end SoC/RAM/Memory.

 

[Alovon11]

2) 8 nm : it hasn't taped out. N6 and 7LPP/5LPE/4LPX are still on the table.

Ehhh
It's likely taped out at this point, at least the rest of the Orin Family outside of maybe ADAS.

Orin AGX is coming out in Q1 2022, and the rest of the Automotive Orin Family with it, so that means Production at the latest has to start in December 2021.

So those are firmly taped out at this point I feel unless they are going to make the quickest turnaround between Tape out to production ever seen.

 

[Skittzo]

5) I would not take Orin having RT cores as an evidence of T239 having RT cores. Nvidia won't need those cores on a shield TV.

The RT stuff comes from Nate too, partially.

 

[Dakhil]

Watch Nintendo Shadow drop the announcement of the Dane at the Game Awards for an early April Launch with BOTW2 getting is name there.

(At the very least we may get Metroid Prime 1 Remake/Remaster revealed at TGA if Emily Rodgers is to be believed on her "Finished Dev over the summer comment. There has to be some reason to hold it for that long, why not to be a title alongside Dane)

Tinfoil Hat off

There probably needs to be reports of mass production for the DLSS model* starting by the end of 2021 at the latest in order for a launch at early 2022 to be realistic. And also nobody knows for 100% certainty if Dane's already taped out or will soon be taped out, especially with NateDrake's comments (here and here) that devkits have limited RTX support, and that RTX support has been shown to have a much more negative impact on battery life than anticipated when testing RTX support for handheld mode. So Nintendo and Nvidia could be making tweaks with regards to RTX support, especially if Nintendo and Nvidia considers RTX support to be very important. And I don't think the RT cores on the Jetson AGX Orin and the Jetson Orin NX are going to be used as extensively as potentially with Dane.

I think Nintendo's gauging people's interest in the Metroid series with Metroid Dread before deciding on the release date for the Metroid Prime remaster/remake.

2) 8 nm : it hasn't taped out. N6 and 7LPP/5LPE/4LPX are still on the table.

5) I would not take Orin having RT cores as an evidence of T239 having RT cores. Nvidia won't need those cores on a shield TV.

2) Although potentially true, the fact that the max CPU and GPU frequencies for the Jetson AGX Orin and the Jetson Orin NX are practically identical to the max CPU and GPU frequencies for the Tegra X1 is a strong indicator that Samsung's 8N process node is used, considering that the CPU and GPU frequencies can definitely at least be set considerably higher if a more advanced process node (Samsung's 7LPP process node or TSMC's N6 process node) is used.

5) NateDrake mentioned that devkits have limited RTX support. And I don't think Nintendo would have included limited RTX support for devkits if Nintendo planned on physically removing the RT cores from the GPU on Dane. Considering that kopite7kimi mentioned that Dane's a custom variant of Orin, I imagine it would be very cost prohibitive to physically remove the RT cores from the GPU on Dane, or redesign the GPU on Dane without RT cores. (Nvidia could simply disable the RT cores for a potential Nvidia Shield TV model equipped with a SoC very similar to Dane.)

 

[LoneRanger]

Don’t want to be pessimistic but expecting a lower node than 8nm on next Switch hardware reminds me when there was (some) speculation that OG Switch was going to have a 16nm SoC.

Things can change of course, but I’m a good believer of 8nm Ampare/Loveface architecture.

about RT, I included because of NateDrake mention. I’m not a 100% tech guy (I’m learning about it ) but if its real, its going to be much more limited than Xbox Series S raytracing. Though its nice if next switch ends up having that and I can see Nintendo using it as a good hardware showcase (with DLSS) on future games like Zelda BOTW2.

 

[Pokemaniac]

5) I would not take Orin having RT cores as an evidence of T239 having RT cores. Nvidia won't need those cores on a shield TV.

Nintendo is the main customer for this chip, not the Nvidia Shield. It's possible something could get added to the chip because an upcoming Shield needs it, but it's not going to be dictating stuff left out.

 

[Alovon11]

Don’t want to be pessimistic but expecting a lower node than 8nm on next Switch hardware reminds me when there was (some) speculation that OG Switch was going to have a 16nm SoC.

Things can change of course, but I’m a good believer of 8nm Ampare/Loveface architecture.

about RT, I included because of NateDrake mention. I’m not a 100% tech guy (I’m learning about it ) but if its real, its going to be much more limited than Xbox Series S raytracing. Though its nice if next switch ends up having that and I can see Nintendo using it as a good hardware showcase (with DLSS) on future games like Zelda BOTW2.

I think the limit is moreso for power consumption.

If they can get it working within the power confines, it would perform very good for the number of cores (Due to NVIDIA's RT cores being over Double the power of the Ray Accelerators in the AMD GPUs)

Like, depending on the game/benchmark, 4RT cores actually would be enough to match the 20 Ray Accelerators in the Series S.

Now in others, it would be 6 or 8, but still, the margin is close.

 

[ILikeFeet]

going back to the ARM demo, we know it's running on their G78, but they don't speak of which configuration. the gpu can go up to 24 clusters, but I can't find a mediatek chip that has that configuration (or even the G78). so it's probably some unreleased chip that's in testing

 

[ILikeFeet]

during Valve's dev-oriented Deck thingie, they shown off some loading comparisons. Nintendo definitely has room to speed up storage

 

[ShadowFox08]

during Valve's dev-oriented Deck thingie, they shown off some loading comparisons. Nintendo definitely has room to speed up storage

Most definitely!

Funny I was reading up on steam deck hardware details from valve released today before coming here, and it's pretty interesting. Definitely worth a read

Valve details Steam Deck hardware, the AMD Zen2/RDNA2 SoC is codenamed Aerith - VideoCardz.com

Steam Deck detailed During Steam Deck Development Live Stream, Valve employees have explained the principles behind the upcoming handheld device. AMD Aerith SoC features 4 cores based on Zen2 microarchitecture with clock speeds varying from 2.4 GHz up to 3.5 GHz. Valve claims that its goal was...

videocardz.com

 

[Alovon11]

Most definitely!

Funny I was reading up on steam deck hardware details from valve released today before coming here, and it's pretty interesting. Definitely worth a read

Valve details Steam Deck hardware, the AMD Zen2/RDNA2 SoC is codenamed Aerith - VideoCardz.com

Steam Deck detailed During Steam Deck Development Live Stream, Valve employees have explained the principles behind the upcoming handheld device. AMD Aerith SoC features 4 cores based on Zen2 microarchitecture with clock speeds varying from 2.4 GHz up to 3.5 GHz. Valve claims that its goal was...

videocardz.com

Very weird how they are pushing what should be limited to 400MB/s up to 500MB/s

is eMMC "Overclockable"? Or does it have to do with them linking it up via PCI-E 2.0?

 

[Look over there]

Hmm, I expected a larger disparity between the eMMC and NVMe. Guess they're maxing out the eMMC and going with a low power active state for the NVMe.

 

[Golden Darkness]

Do we know the size of the Steam Deck SoC? If not, can we estimate its size? We do know it has a Zen 2 4 Core CPU/8 Threads and 8 RDNA2 Compute Units.

 

[Dekuman]

Again 8SMs is the most likely config cost-wise at this point due to Orin NX being right there as Z0m3le has pointed out on a few occasions.

And while 4RT cores may not sound like a lot there are some things to consider.

• NVIDIA's RT Cores are far ahead of the Ray Accelerators from AMD.
  • Assuming worst-case for NVIDIA vs AMD RT (Which likely isn't the case as Turing RT cores are stronger than Ray Accelerators) You'd only need 10 RT cores to Match the Series S's 20 Ray Accelerators.
    • And like I mentioned earlier, Turing is actually better at RT than RDNA2 in most benchmarks/tests, so that number is actually smaller to match the Series S's 20.
      • Some number configurations actually put 4RT cores as matching the 20RA's but still trying to number-crunch to verify that, but still it may be close.
• RT has gotten far more scalable
  • RTXGI (NVIDIA's own GI solution) is scalable enough to run in software on the OG Xbox One, and considering Dane when docked will be stronger than the OG PS4, those RT cores can be used to accelerate RTXGI.
    • Same thing with the RT cores accelerating things that are in software like Lumen or CryTek's RT.
  • RT Shadows are able to be run on smartphone ARM SoC GPUs in software now (With crap Mali drivers), so they can be accelerated perfectly fine to run on Switch on a low number of RT cores.
  • And even if not RT shadows, there is RTAO, RT Audio, and other usages of RT that can be used (Maybe even acceleration of SSRT effects like SSR, SSRTGI.etc)
• DLSS
  • Back to the Series S point, even assuming Dane has half the theoretical RT performance of the Series S, Dane can easily run at half the internal resolution or even a quarter of the resolution and get that performance back and then some and jump to match or surpass the Series S in output (outside of the resolution of the RT effect suffering, but for RTGI, RTAO, and RT Shadows that shouldn't be too noticeable)

I'm just being a pessimist. But stepping away from technology side and just looking at this from a 'next console' checkbox from a practical purpose, current Switch is not really hitting 1080p native docked and 720p portable very often and usually at 30fps only. So I assume they will want a device that is bare minimum 3x-4x more powerful before DLSS, with i think a baseline 1-2x over Switch dedicated just to hitting 720p/1080p rendering at a solid 30 fps in portable/docked. and the remaining power to increase geometry/effects/fps. Total package being 3-4x.

So I was thinking 2x the CUDA cores at roughly 2x the clocks for the GPU as the upper range. And if clocks can't reasonably be doubled due to heat, then they will need more cores to compensate. is that a reasonable estimate?

 

[Dakhil]

I was pondering about Thraktor's speculation that Orin X's a different binning of the die used for the Jetson AGX Orin and the Jetson Orin NX, which is how Orin X can achieve 254 TOPS, considering the max CPU and GPU frequencies for the Jetson AGX Orin and the Jetson Orin NX are practically identical to the max CPU and GPU frequencies for the Tegra X1. And I imagine achieving 254 TOPS could be quite difficult, especially if Samsung's 8N process node is used.

And then, I remembered seeing Li Auto's press release mention that Orin was fabricated using a 7 nm** process node back in 27 September 2020.
However, when I checked the official press release on Li Auto's website, Li Auto mentioned that Orin was fabricated using a 8 nm** process node. But when I took a look at Li Auto's press release at GlobeNewswire, the official press release mentioned that Orin was fabricated using a 7 nm** process node. (This definitely seems very suspicious to me.)

So more proof that Orin's fabricated using Samsung's 8N process node? And more proof that IM Motors could have been speculating that Orin X's been fabricated using a 7 nm** process node?

** → a marketing nomenclature used by all foundry companies

Do we know the size of the Steam Deck SoC? If not, can we estimate its size? We do know it has a Zen 2 4 Core CPU/8 Threads and 8 RDNA2 Compute Units.

According to Patrick Schur on Twitter, Van Gogh is around ~162.61 mm². (Patrick Schur seems to be correct on the Van Gogh details.)

 

[ILikeFeet]

So I was thinking 2x the CUDA cores at roughly 2x the clocks for the GPU as the upper range. And if clocks can't reasonably be doubled due to heat, then they will need more cores to compensate. is that a reasonable estimate?

from the looks of things, we aren't getting double clocks. we might get around the Switch's theoretical peaks (920MHz), however. the issue with 2X the cuda cores is that it's questionable how much RT acceleration is useable. we know it "works", but enough to provide a meaningful gain without cranking up heat?

 

[Alovon11]

I'm just being a pessimist. But stepping away from technology side and just looking at this from a 'next console' checkbox from a practical purpose, current Switch is not really hitting 1080p native docked and 720p portable very often and usually at 30fps only. So I assume they will want a device that is bare minimum 3x-4x more powerful before DLSS, with i think a baseline 1-2x over Switch dedicated just to hitting 720p/1080p rendering at a solid 30 fps in portable/docked. and the remaining power to increase geometry/effects/fps. Total package being 3-4x.

So I was thinking 2x the CUDA cores at roughly 2x the clocks for the GPU as the upper range. And if clocks can't reasonably be doubled due to heat, then they will need more cores to compensate. is that a reasonable estimate?

Well, you can't really look at it that way though.

If Nintendo were completely designing this custom (or off Xaiver), maybe.

Heck, if the idea that the Mariko models were to have a power-boost was true then what you suggested likely would've been the case (Locking Native Res output targets).

But we aren't in that scenario and Nintendo is sort of bound to NVIDIA's progress in ARM, and that is Orin, specifically Orin NX to base Dane on.

 

[Dekuman]

Well, you can't really look at it that way though.

If Nintendo were completely designing this custom (or off Xaiver), maybe.

Heck, if the idea that the Mariko models were to have a power-boost was true then what you suggested likely would've been the case (Locking Native Res output targets).

But we aren't in that scenario and Nintendo is sort of bound to NVIDIA's progress in ARM, and that is Orin, specifically Orin NX to base Dane on.

If that's that case what if nvidia can't deliver even those basic requirements? I think a succ being less than 3-4x more powerful than the previous device would be unacceptable IMHO.

I am looking at it from a Nintendo requirements perspective, which I assume nvidia would have on hand when desinging the SoC.

Edit: you seem fairly certian 1024 (8 SM?) cores is not unreasonable for the new Switch. how big is going the die going to end up being? compared to OG Switch. I understand we're going from 20 nm to 8nm here, but since the scale is largely marketing these days, would 8nm allow for 1056 cores at the same die size?

 

[NineTailSage]

1) Ampere architecture : actually it could be Lovelace/Ampere/Turing or a custom version that will probably be named after the last revealed Desktop Architecture. If it is revealed in H2 2022, I would expect Nvidia to call it Lovelace even if it is 200% Pascal.

2) 8 nm : it hasn't taped out. N6 and 7LPP/5LPE/4LPX are still on the table.

5) I would not take Orin having RT cores as an evidence of T239 having RT cores. Nvidia won't need those cores on a shield TV.

6) They could target the Mariko 10W power budget especially if they are planning to release lite at the same time.

1) IMHO the console form factor was probably the OLED form factor before the chip shortages cutting Nintendo business case before signing the contract for the high end SoC/RAM/Memory.

RT cores aren't exclusive to graphical features, Nvidia could use them in a new Shield TV for processing 3D audio and sell that as a high end feature for audiophiles.

 

[Look over there]

Hmm, I expected a larger disparity between the eMMC and NVMe. Guess they're maxing out the eMMC and going with a low power active state for the NVMe.

Continuing this train of thought: assuming that eMMC 5.1 hits 400 MB/s, then the NVMe drive is probably using an active state of about 1 W, assuming average-ish power efficiency.
On the other hand, if the eMMC is hitting 250 MB/s, then that NVMe drive is either shooting for some fraction of a watt, or it's 1 W of some shittastic efficiency.

...oh, right, there's an SD Card there in a graph. Uh... 100 MB/s max, right? It is UHS-I, right? And that only takes 18% more time to load a game than the 512 gb SSD? Or am I forgetting something here?

 

[Z0m3le]

We don't know what kind of RT support the Dane SoC will have, but it will likely have some support now.
The big news as I understand it this week is it was revealedthe Orin design has RT cores (1 per 2 SM) so maybe we'll get 2-3 RT cores in the final Nintendo Switch 2 SoC, which isn't a lot, but it is better than 0 we were expecting back earlier this year.

I'd suggest 4RT Cores are very possible, as 8SM is the most likely configuration at the moment and Orin has 1RT:2SM, however Orin isn't designed for gaming and is using RT cores in much more limited capacity. Dane is a gaming SoC and could easily add back the RT cores missing from traditional Ampere architecture, there is other logic on a traditional GPU that a gaming device doesn't need and Orin has a bunch of accelerators that don't make sense to keep as well.

Honestly, the number of RT cores to me is either 0, 4, 6, or 8. 0 if they need to remove the RT cores to make room, 4 if they don't change the configuration from Orin, 6 if they go with 6SM and 8RT if raytracing is a goal for Dane, they will put back the RT cores found in Ampere architecture.

The main point is Orin isn't designed for games, having 8 RT cores is a waste of space, but the Jetson devkits are designed for experimental codes, discovery and driving the AI field, it's not made for game development.

 

[Alovon11]

If that's that case what if nvidia can't deliver even those basic requirements? I think a succ being less than 3-4x more powerful than the previous device would be unacceptable IMHO.

I am looking at it from a Nintendo requirements perspective, which I assume nvidia would have on hand when desinging the SoC.

Edit: you seem fairly certian 1024 (8 SM?) cores is not unreasonable for the new Switch. how big is going the die going to end up being? compared to OG Switch. I understand we're going from 20 nm to 8nm here, but since the scale is largely marketing these days, would 8nm allow for 1056 cores at the same die size?

The thing is that it isn't limited to TX1's die size, especially if it's a 400$ product.

Heck, I bet you that it will be around or smaller than the Steam Deck's 160mm Van Gogh APU with 8 A78Cs and 8SMs.

ARM is far more power efficient than x86, and the A78Cs are far smaller than Zen2 CPU Cores.

 

[Deleted member 645]

See below

 

[Deleted member 645]

Don’t want to be pessimistic but expecting a lower node than 8nm on next Switch hardware reminds me when there was (some) speculation that OG Switch was going to have a 16nm SoC.

Things can change of course, but I’m a good believer of 8nm Ampare/Loveface architecture.

about RT, I included because of NateDrake mention. I’m not a 100% tech guy (I’m learning about it ) but if its real, its going to be much more limited than Xbox Series S raytracing. Though its nice if next switch ends up having that and I can see Nintendo using it as a good hardware showcase (with DLSS) on future games like Zelda BOTW2.

Yep. I mean I’ve learned my lesson “many” times over. I’m expecting the lowest case most cautious scenario. There are so many factors in play and things we don’t know yet. It’s hard for me to say for sure in anything.

 

[Z0m3le]

Something we also aren't discussing is any modifications Nintendo might make, they do have engineers that could work with Nvidia on specific GPU features, they could modify RT cores, Tensor cores, Cuda Cores (which will likely be modified for Maxwell BC anyways) or other components to have some unique features that Nintendo wants to focus on. This could be as simple as putting back RT cores that Orin doesn't have, or could modify tensor cores to do other things... There is a million different things they could do.

 

[Dekuman]

Something we also aren't discussing is any modifications Nintendo might make, they do have engineers that could work with Nvidia on specific GPU features, they could modify RT cores, Tensor cores, Cuda Cores (which will likely be modified for Maxwell BC anyways) or other components to have some unique features that Nintendo wants to focus on. This could be as simple as putting back RT cores that Orin doesn't have, or could modify tensor cores to do other things... There is a million different things they could do.

if we take the speculation (?) that Ninteod is looking to build its only DLSS solution with technology they own, what sort of modifications could we see? I assume the software AI learning side is likely to build on nvidia's expertise with DLSS but branded/owned by Nintendo, but on the hardware side would anything really change?

 

[NineTailSage]

There probably needs to be reports of mass production for the DLSS model* starting by the end of 2021 at the latest in order for a launch at early 2022 to be realistic. And also nobody knows for 100% certainty if Dane's already taped out or will soon be taped out, especially with NateDrake's comments (here and here) that devkits have limited RTX support, and that RTX support has been shown to have a much more negative impact on battery life than anticipated when testing RTX support for handheld mode. So Nintendo and Nvidia could be making tweaks with regards to RTX support, especially if Nintendo and Nvidia considers RTX support to be very important. And I don't think the RT cores on the Jetson AGX Orin and the Jetson Orin NX are going to be used as extensively as potentially with Dane.

I think Nintendo's gauging people's interest in the Metroid series with Metroid Dread before deciding on the release date for the Metroid Prime remaster/remake.


2) Although potentially true, the fact that the max CPU and GPU frequencies for the Jetson AGX Orin and the Jetson Orin NX are practically identical to the max CPU and GPU frequencies for the Tegra X1 is a strong indicator that Samsung's 8N process node is used, considering that the CPU and GPU frequencies can definitely at least be set considerably higher if a more advanced process node (Samsung's 7LPP process node or TSMC's N6 process node) is used.

5) NateDrake mentioned that devkits have limited RTX support. And I don't think Nintendo would have included limited RTX support for devkits if Nintendo planned on physically removing the RT cores from the GPU on Dane. Considering that kopite7kimi mentioned that Dane's a custom variant of Orin, I imagine it would be very cost prohibitive to physically remove the RT cores from the GPU on Dane, or redesign the GPU on Dane without RT cores. (Nvidia could simply disable the RT cores for a potential Nvidia Shield TV model equipped with a SoC very similar to Dane.)

Nvidia are definitely more aware of what Ampere is good at now and how far the architecture can be pushed, the new A2 Tensor gpu is 40-60 watt based on GA107 and is also passively cooled at that. I think a modified or under clocked version of GA107 is very possible as the development hardware used in early dev-kits, as the chip is probably the base for Orin's overall design...

Also I think Nintendo are just giving enough breathing room between Metroid Dread and the announcement of a Metroid Prime remaster, as to not cannibalize the game from selling.

NVIDIA Launches A2 Tensor Core GPU, An Entry-Level Design Powered By Ampere GA107 GPU & 16 GB GDDR6 Memory

NVIDIA has further expanded its professional data center lineup of Ampere GPUs with the A2 Tensor Core GPU accelerator.

wccftech.com

 

[ILikeFeet]

Yep. I mean I’ve learned my lesson “many” times over. I’m expecting the lowest case most cautious scenario. There are so many factors in play and things we don’t know yet. It’s hard for me to say for sure in anything.

get excited for a 2SM, 1RT core, 4 A78 Switch 2 then! only 50% performance uplift! for only $400

 

[Skittzo]

if we take the speculation (?) that Ninteod is looking to build its only DLSS solution with technology they own, what sort of modifications could we see? I assume the software AI learning side is likely to build on nvidia's expertise with DLSS but branded/owned by Nintendo, but on the hardware side would anything really change?

Tangent but I don't think they plan to use their own ML upscaling solution for new games, but rather to upscale their back catalog.

 

[NineTailSage]

The best thing that the A2 Tensor gpu gives us is a spectrum of where the full Orin graphics chip would be without the DLA hardware.
Of course they still don't divulge the die size, even though Kopite has hinted it being under 200mm²(around 19*mm²).

NVIDIA introduces A2 Tensor Core GPU with GA107 GPU and 16GB GDDR6 memory - VideoCardz.com

NVIDIA A2 is a low-power accelerator for AI Today NVIDIA launches its entry-level edge interference accelerator. NVIDIA continues to add more SKUs to its Data Center Ampere lineup. The series which we used to call “Tesla” are now using a much simpler naming, albeit a lot more confusing. The A2...

videocardz.com

 

[Alovon11]

The best thing that the A2 Tensor gpu gives us is a spectrum of where the full Orin graphics chip would be without the DLA hardware.
Of course they still don't divulge the die size, even though Kopite has hinted it being under 200mm²(around 19*mm²).

NVIDIA introduces A2 Tensor Core GPU with GA107 GPU and 16GB GDDR6 memory - VideoCardz.com

NVIDIA A2 is a low-power accelerator for AI Today NVIDIA launches its entry-level edge interference accelerator. NVIDIA continues to add more SKUs to its Data Center Ampere lineup. The series which we used to call “Tesla” are now using a much simpler naming, albeit a lot more confusing. The A2...

videocardz.com

If so, that would mean the GPU would be indeed less than 100mm^2 in Dane, and maybe even smaller if they keep the halved RT Core count.

 

[ILikeFeet]

The best thing that the A2 Tensor gpu gives us is a spectrum of where the full Orin graphics chip would be without the DLA hardware.
Of course they still don't divulge the die size, even though Kopite has hinted it being under 200mm²(around 19*mm²).

NVIDIA introduces A2 Tensor Core GPU with GA107 GPU and 16GB GDDR6 memory - VideoCardz.com

NVIDIA A2 is a low-power accelerator for AI Today NVIDIA launches its entry-level edge interference accelerator. NVIDIA continues to add more SKUs to its Data Center Ampere lineup. The series which we used to call “Tesla” are now using a much simpler naming, albeit a lot more confusing. The A2...

videocardz.com

it's probably a cut down GA107

 

[Alovon11]

So, putting this here as it would pretty much be how 2TFLOP Dane would perform before Console Optimization/DLSS (2TFLOP Dane would be around the power of a GTX 1050Ti, at least assuming my 30% improvement over Ampere due to the Cache lowball)

 

[ReddDreadtheLead]

I'm curious on the specint2006 being 240, being an improvement over the 137 of Xavier.

if you are wondering where I got it from, it is from the white papers:

https://www.nvidia.com/content/dam/en-zz/Solutions/gtcf21/jetson-orin/nvidia-jetson-agx-orin-technical-brief.pdf

it seems like a mostly useless thing to bring up.

 

[ILikeFeet]

I'm curious on the specint2006 being 240, being an improvement over the 137 of Xavier.

if you are wondering where I got it from, it is from the white papers:

https://www.nvidia.com/content/dam/en-zz/Solutions/gtcf21/jetson-orin/nvidia-jetson-agx-orin-technical-brief.pdf

it seems like a mostly useless thing to bring up.

Specint measures general computing performance. It's a bunch of tasks one might do on a PC. Not really useless as a dev board, but for a gaming device, it isn't representative of performance

 

[ReddDreadtheLead]

Specint measures general computing performance. It's a bunch of tasks one might do on a PC. Not really useless as a dev board, but for a gaming device, it isn't representative of performance

I thought the move was to 2017 over the 2006 one?

 

[ILikeFeet]

I thought the move was to 2017 over the 2006 one?

Nvidia is probably using the old one because it makes them look better

 

[Mr Swine]

Does anyone know how much bandwidth DLSS used to scale up to 4K compared to true 4K resolution? Does using AI to scale up 1080p image to 4K use more bandwidth than 1080p but far less than 4K?

 

[Dekuman]

So, putting this here as it would pretty much be how 2TFLOP Dane would perform before Console Optimization/DLSS (2TFLOP Dane would be around the power of a GTX 1050Ti, at least assuming my 30% improvement over Ampere due to the Cache lowball)

2TF docked perf before dlss would be roughly 5x performance of the current Switch. I recall prior speculation was closer to 1.2 TF (this was many months ago) what has changed to push us closer to the 2TF benchmark? is it only because it's not going to be Pro device like we were speulating previously?

 

[ILikeFeet]

given the Steam Deck is around the 1050 based on some fairly old videos, that's my expectation. Docked Dane and unbridled Steam Deck are gonna trade blows in non-cpu limited situations, I think (sans dlss)

 

[Alovon11]

2TF docked perf before dlss would be roughly 5x performance of the current Switch. I recall prior speculation was closer to 1.2 TF (this was many months ago) what has changed to push us closer to the 2TF benchmark? is it only because it's not going to be Pro device like we were speulating previously?

Not all TFLOPs are the same.

But the 2TFLOP GPU Number is pretty much what Orin NX is, and that is with the thermal and power loss from the A78AEs, DLAs, and the turned off components that are sitting there.

So while Orin NX is rated for 25W peak, that is likely because the A78AEs are more power hungry, the DLAs, and other Automotive components.

So getting close to that should be reasonable for a 15W TDP if they lower CPU Clocks after removing those

 

[Look over there]

The SPECInt 2k6 comparison sounds reasonably close to ARM's projections for A75 (Carmel's A75-ish in perf)->A76->A77->A78, with the remaining difference probably due to a slight difference in clocks.
Btw, as I remember it, A75->A76 and A76->A77 were projected to have somewhat higher gains in floating point than integer.

 

[NineTailSage]

My question is if Nvidia are willing to use a passive design for the A2(which yes is a data center gpu that runs between 40-60 watts).
What are the chances Nintendo either goes with something slightly higher than what we are expecting max powered while docked or a (7-15watt handheld to docked design) that is passively cooled as well?

 

[Mercury_Sagit]

My question is if Nvidia are willing to use a passive design for the A2(which yes is a data center gpu that runs between 40-60 watts).
What are the chances Nintendo either goes with something slightly higher than what we are expecting max powered while docked or a (7-15watt handheld to docked design) that is passively cooled as well?

I guess going from an actively cooling design to a passive one will be a trade off between space for fans and heat fins. Nintendo can go for the latter if it chooses the wide and slow hardware config. But still, I don't really see the benefit of such design, when the downside can be visible in most of the smartphone and tablet models nowadays, since they tend to throttle under high load.

 

[Terrell]

My question is if Nvidia are willing to use a passive design for the A2(which yes is a data center gpu that runs between 40-60 watts).
What are the chances Nintendo either goes with something slightly higher than what we are expecting max powered while docked or a (7-15watt handheld to docked design) that is passively cooled as well?

I've been toying with this idea of Nintendo going to 20W TDP while docked, because I think that's the highest they could hit without compromising the hybrid design and having the system run too hot, but I can't see it going any higher, and especially not with passive cooling design.

Passive cooling for a hybrid device would require them to entirely re-think how heat management is done with the device. With the hardware being used as it is (sustained higher load in a non-stationary device), I'm fairly certain that passive cooling isn't going to be an option they can pursue without a drastic reduction in its docked performance to compensate.

 

[ILikeFeet]

My question is if Nvidia are willing to use a passive design for the A2(which yes is a data center gpu that runs between 40-60 watts).
What are the chances Nintendo either goes with something slightly higher than what we are expecting max powered while docked or a (7-15watt handheld to docked design) that is passively cooled as well?

there's no reason not to do a passively cooled design. hell, even phones are coming up with crazy ways to cool themselves within the space constraints. not to mention, it would need more exposed portions to dissipate heat and probably a dock designed to move air (even if passively). that's a lot of R&D for something they know works very well

 

[NineTailSage]

I guess going from an actively cooling design to a passive one will be a trade off between space for fans and heat fins. Nintendo can go for the latter if it chooses the wide and slow hardware config. But still, I don't really see the benefit of such design, when the downside can be visible in most of the smartphone and tablet models nowadays, since they tend to throttle under high load.

Phones definitely have throttling issues because they are also clocked much higher than Switch cores will ever be.

I've been toying with this idea of Nintendo going to 20W TDP while docked, because I think that's the highest they could hit without compromising the hybrid design and having the system run too hot, but I can't see it going any higher, and especially not with passive cooling design.

Passive cooling for a hybrid device would require them to entirely re-think how heat management is done with the device. With the hardware being used as it is (sustained higher load in a non-stationary device), I'm fairly certain that passive cooling isn't going to be an option they can pursue without a drastic reduction in its docked performance to compensate.

The idea that we could see a redesign to the Switch form factor that allows it to be passively cooled in handheld mode(because it's only 6-8 watts) and then the dock actually has a forced air design while that pushes air through the bottom out the top while in docked mode.
Putting the fan in the dock would definitely give more real-estate for better components and specs than the current Switch's MB allows for.

there's no reason not to do a passively cooled design. hell, even phones are coming up with crazy ways to cool themselves within the space constraints. not to mention, it would need more exposed portions to dissipate heat and probably a dock designed to move air (even if passively). that's a lot of R&D for something they know works very well

It's definitely a possibility and not having a fan constantly running has to save on battery power, I can see them at least testing out the what if scenario of all of this and that A2 gpu was just the spark of huh maybe....
Even looking at the venting on the card, it's extremely compact and has openings on the front and back that's it...

 

[Terrell]

there's no reason not to do a passively cooled design. hell, even phones are coming up with crazy ways to cool themselves within the space constraints. not to mention, it would need more exposed portions to dissipate heat and probably a dock designed to move air (even if passively). that's a lot of R&D for something they know works very well

The problem is right in the video, with the temperature being 47.7C. The current Switch at a 15W TDP handily hits that temperature or slightly higher with a slight over-abundance of active cooling in place; no passive cooling system is efficient enough to achieve better results than that, especially for a system that could be removed from a stationary position with air pushed through it at any moment.