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

[BlackTangMaster]

From the Jetson AGX Orin data sheet (v0.2). Regarding its GPU Architecture:

Any indication on how many GPU cores are available at 15W in the data sheet? And the clocks used @15W?

 

[ILikeFeet]

just logged into my dev account to check for myself. weird that Jensen never mentioned it on the stream

Any indication on how many GPU cores are available at 15W in the data sheet? And the clocks used @15W?

nope

 

[fwd-bwd]

Any indication on how many GPU cores are available at 15W in the data sheet? And the clocks used @15W?

Maybe I missed it but so far I couldn't find any info on the 15w mode. Note that this data sheet is marked "preliminary".

 

[ILikeFeet]

all GPU cores will be available in all power mode since you can't turn off gpu cores without fusing them off. Xavier was like this

 

[Whatnoww]

I also agree that "Switch 4k" name is also too loaded and Nintendo are the ones to flaunting technical specs. They also won't be able to consistently deliver 4k resolution. Maybe for current switch games, but for Xbox series s ports, let alone ps4 tier games? Definitely not. 2k is the most we'll likely see for PS4 level quality (basically ps4 pro level or slightly more )


Remember also that DLSS needs a lot of GPU power/tensor cores to go from 1080p to 4k and have it refreshed in a reasonable amount of milliseconds. We're talking about DLSS support for a GPU as powerful as a PS4 base, not a RTX 2060 or more.2060 is more than 3z as powerful as a PS4 in GPU power.

So maybe 1080p to 2k at most for PS4 level quality games.

Why would increasing the vertical width of the picture by a few pixels and changing the aspect ratio do anything for Nintendo? I don't think they'll be moving from 16:9 any time soon. They'll stick to 1080p 16:9 and I suspect most developers will offer a performance mode that will run at 1080p. Rendering 100 extra pixels on the horizontal axis is not going to impress anyone.

 

[fwd-bwd]

Regarding the Jetson AGX Orin's power modes, from the Technical Brief (emphases mine):

"Jetson AGX Orin supports three optimized power budgets: 15W, 30W, and 50W. Each power mode caps various component frequencies, and the number of online CPU, GPU TPC, DLA, and PVA cores. Customers can leverage the nvpmodel in Jetson Linux to use one of these pre-optimized power modes or to customize their own power mode within the design constraints provided in our documentation."

For those hoping for a UFS storage, the Jetson AGX Orin support the single lane UFS. It's not as fast as the dual lane UFS but still pretty good.

 

[ILikeFeet]

the question now is what kind of RT performance can be had with a Dane Switch setup. would definitely make stuff like Lumen and RTXGI and RTXDDI faster

 

[Hermii]

We all though Nintendo would need to do customisations to include RT. Turns out, they actually need customisations to remove it.

 

[Mr Swine]

Jetson AGX Orin includes RT cores its-happening.gif

Honestly I don’t see Switch using the RT cores if not for docked mode only running at a much higher power consumption (20+w?) combined with DLSS

 

[BlackTangMaster]

all GPU cores will be available in all power mode since you can't turn off gpu cores without fusing them off. Xavier was like this

Xavier was able to cut CPU and TPC. It was even able to cut the DVA and PLA.

Regarding the Jetson AGX Orin's power modes, from the Technical Brief (emphases mine):

"Jetson AGX Orin supports three optimized power budgets: 15W, 30W, and 50W. Each power mode caps various component frequencies, and the number of online CPU, GPU TPC, DLA, and PVA cores. Customers can leverage the nvpmodel in Jetson Linux to use one of these pre-optimized power modes or to customize their own power mode within the design constraints provided in our documentation."

For those hoping for a UFS storage, the Jetson AGX Orin support the single lane UFS. It's not as fast as the dual lane UFS but still pretty good.

Just like Xavier that was able to cut the number of CPU and GPU cores for less power consumption.

 

[ILikeFeet]

Xavier was able to cut CPU and TPC. It was even able to cut the DVA and PLA.


Just like Xavier that was able to cut the number of CPU and GPU cores for less power consumption.

looking it up again, we're both right. the Jetson Xavier NX keeps all TPCs. Jetson Xavier AGX has modular number of TPC activations

Welcome — Jetson Linux<br/>Developer Guide 34.1 documentation

docs.nvidia.com

 

[Mercury_Sagit]

Looking at the Orin SoC techinical brief, I actually see DLA (deep learning accelerator) as a viable component to keep in Dane, just to load it with solely the neural net for DLSS.

Earlier this year, both Nintendo and NV had patent/research mentioning embedding a whole neural network into the GPU and eliminate VRAM IO by optimizing the neural net structure to match the GPU's cache structure. So if these two optimize DLSS in this direction, DLA can stay for DLSS, leaving the tensor core to do other tasks: ML for games' AI, RTX IO (?), etc. Another supporting argument is that DLA is a fixed-function hardware, which generally has smaller die area, memory, power usage i.e. better efficiency. A counter argument is of course whether NV and Nintendo can actually fit DLSS into DLA without hampering the output quality.

 

[BlackTangMaster]

Looking at the Orin SoC techinical brief, I actually see DLA (deep learning accelerator) as a viable component to keep in Dane, just to load it with solely the neural net for DLSS.

Earlier this year, both Nintendo and NV had patent/research mentioning embedding a whole neural network into the GPU and eliminate VRAM IO by optimizing the neural net structure to match the GPU's cache structure. So if these two optimize DLSS in this direction, DLA can stay for DLSS, leaving the tensor core to do other tasks: ML for games' AI, RTX IO (?), etc. Another supporting argument is that DLA is a fixed-function hardware, which generally has smaller die area, memory, power usage i.e. better efficiency. A counter argument is of course whether NV and Nintendo can actually fit DLSS into DLA without hampering the output quality.

The DLA being outside the GPU would induce too much latency.

 

[Alovon11]

The DLA being outside the GPU would induce too much latency.

On other ideas, could the DLA act as a sort of ESRAM/L3 Cache replacer?

At least an accelerator of sort (sort of how Nintendo speeds loading times by boosting the CPU clock in the Switch RN)

 

[Dakhil]

Yeah, probably for power savings only. sigh. I'm guessing lpddr5x is a q4 release then. Hard to believe lpddr5 is 3 years old now. There is some speculation it could be available as soon as on the Samsung Galaxy S22 phone models though, which will be out in February I believe.

Andrei Frumusanu from Anandtech believes that LPDDR5X will become available for 2023 devices.

And I highly doubt it since Samsung mentioned starting development of LPDDR5X DRAM, not starting mass production of LPDRR5X DRAM.

~

Anyway, in terms of video encode streams, AGX Orin supports H.264, H.265 (HEVC), and AV1.

 

[ILikeFeet]

I can see nintendo going with VP9 for video encoding

 

[Dekuman]

Honestly I don’t see Switch using the RT cores if not for docked mode only running at a much higher power consumption (20+w?) combined with DLSS

Said this before the gap between docked perf and portable perf on the current Switch is a bit too narrow, there's no need to power consumption at 15w in docked mode. Actually I think the Switch's overall power draw is at 15w (including charging the battery/ joycons) so the actual SoC is around 12w (?) docked. If they aim for 15w docked SoC, it would give them more headroom for docked mode.

 

[Dakhil]

So Nvidia also mentioned the Jetson Orin NX, which has 8 Cortex-A78AE cores running at a max frequency of 2 GHz and a GPU based on Ampere with 1024 CUDA cores and 32 Tensor cores (and 8 RT cores) running at a max frequency of 1 GHz.

 

[TonyBaduy]

About the Dane Switch, if they really reduce the amount of cores, people here were speculating it'd be half of AGX or X, which seem to be 12 cores and the Dane would end up at 6, but if the configuration is clusters of 4, wouldn't that mean they'd need to go with either 4 or 8? I doubt they'd cut it to 1/3, so I guess it'd be 8 cores, right?

 

[BlackTangMaster]

So Nvidia also mentioned the Jetson Orin NX, which has 8 Cortex-A78AE cores running at a max frequency of 2 GHz and a GPU based on Ampere with 1024 CUDA cores and 32 Tensor cores (and 8 RT cores) running at a max frequency of 1 GHz.

2TFLOPs@25W is high but it is roughly half the full chip power consumption at 50W.

 

[Skittzo]

About the Dane Switch, if they really reduce the amount of cores, people here were speculating it'd be half of AGX or X, which seem to be 12 cores and the Dane would end up at 6, but if the configuration is clusters of 4, wouldn't that mean they'd need to go with either 4 or 8? I doubt they'd cut it to 1/3, so I guess it'd be 8 cores, right?

I think the idea is they would actually go with different A78 cores. A78AE isn't as useful for gaming as A78C or something similar.

 

[Dakhil]

About the Dane Switch, if they really reduce the amount of cores, people here were speculating it'd be half of AGX or X, which seem to be 12 cores and the Dane would end up at 6, but if the configuration is clusters of 4, wouldn't that mean they'd need to go with either 4 or 8? I doubt they'd cut it to 1/3, so I guess it'd be 8 cores, right?

I think it ultimately depends on what Nintendo wants, considering that the Cortex-A78C allows the option for the hexa-core (6) configuration or the octa-core configuration (8) of the Cortex-A78C, which uses Cortex-A78 cores. I think the octa-core configuration of the Cortex-A78C is likely, considering Jetson Orin NX has 8 Cortex-A78AE cores.

 

[BlackTangMaster]

About the Dane Switch, if they really reduce the amount of cores, people here were speculating it'd be half of AGX or X, which seem to be 12 cores and the Dane would end up at 6, but if the configuration is clusters of 4, wouldn't that mean they'd need to go with either 4 or 8? I doubt they'd cut it to 1/3, so I guess it'd be 8 cores, right?

Dane could use A78C in a 6 or 8 cores configuration in one cluster instead of standard A78/A78AE.

 

[Mercury_Sagit]

Said this before the gap between docked perf and portable perf on the current Switch is a bit too narrow, there's no need to power consumption at 15w in docked mode. Actually I think the Switch's overall power draw is at 15w (including charging the battery/ joycons) so the actual SoC is around 12w (?) docked. If they aim for 15w docked SoC, it would give them more headroom for docked mode.

I think a potential issue with a more power-hungry SoC is thermal, especially in docked mode. Right now Switch draws up to 18W in handheld mode, but the Mariko SoC only consumes ~8W from that. In docked mode, there is an article from Anandtech which shows Erista's energy consumption to be around 11W.
For a (up to) 15W SoC to work without overheating, Nintendo will have to redesign the cooling system or regulate (much) less energy into charging the internal battery when Dane is active.
Edit:

Orin NX's power budget is between 10w and 25w.

The hardware profile for Orin NX at 10W would be a very interesting pointer to possible setup for Dane I think, since Erista and Mariko's energy usage is in the same ballpark in docked mode.

 

[fwd-bwd]

Orin NX's power budget is between 10w and 25w.

 

[Eslayer]

updated roadmap

 

[Skittzo]

Why are there so many NXes? So confusing.

 

[BlackTangMaster]

updated roadmap

There is a hole between Orin NX and nano Next.

 

[Deleted member 645]

You guys are really on top of this shit. Top level investigators.

 

[ILikeFeet]

interesting to note, that Orin has 1 RT Core per 2 SMs. desktop Ampere has 1 RT Core per 1 SM

There are multiple texture processing clusters (TPC) units within a graphics processing cluster (GPC), each TPC includes two SMs, a Polymorph Engine, two Texture Units, and a Ray Tracing core (RTcore)

 

[BlackTangMaster]

interesting to note, that Orin has 1 RT Core per 2 SMs. desktop Ampere has 1 RT Core per 1 SM

Good bye Lighting RT.

Hello audio RT.

 

[ILikeFeet]

Good bye Lighting RT.

Hello audio RT.

don't count it out so quickly. it could still accelerate RTXGI, which is already a software solution, and UE5's Lumen. both are probe-based lighting solutions. not even mobile will support Lumen

 

[LoneRanger]

Orin NX with Switch clock’s (307/384/768 MHz for GPU and 1GHz for CPU) would be a 628/786/1572 GFlops performance. Its 4 times actual Switch (+ DLSS). On docked it’s basically a PS4 with RT support.

CPU (if its using 7 cores for games) would be a ~7-8 jump. Thats very good and honestly, I don’t see it insane. New 3DS was a 6x jump. That x2 or x3 PS4/XOne CPU?

RAM should be 8GB, with less than 1GB for OS, its double than actual hardware.

As an iterative successor aiming for the next 18 months, its specs would be good. It has all that Nintendo needs to display their next wave of projects at full 4k and gives third party developers the chance to port all their stuff that Switch couldn’t make it.

 

[Sol]

So Nvidia also mentioned the Jetson Orin NX, which has 8 Cortex-A78AE cores running at a max frequency of 2 GHz and a GPU based on Ampere with 1024 CUDA cores and 32 Tensor cores (and 8 RT cores) running at a max frequency of 1 GHz.

This seems closer to what the next Switch would be based on, right?

Orin "NX" is also an interesting name...

 

[lemonfresh]

So dane switch is going to be a beast of a device?

 

[Skittzo]

So dane switch is going to be a beast of a device?

It has the potential to be. The specs being discussed here are imo the best case scenario, it's likely to be less impressive than those. But still a good upgrade.

 

[Alovon11]

So dane switch is going to be a beast of a device?

Yeah, as the minimum floor tech-wise has risen a lot.

 

[BlackTangMaster]

So dane switch is going to be a beast of a device?

Nvidia will have to put Orin NX Perfs in a smaller chipset aiming for the original TX1 SoC power consumption which is not guaranteed. NX 15W power profile will probably be close to what we could expect from a custom Nintendo chip.

We know it would be either less than 1GHz or 1024 cuda cores maybe both. The CPU is still unknown as it could use a customised 2GHz 8*A78C one. But overall, it would be less than that.

 

[Pokemaniac]

This seems closer to what the next Switch would be based on, right?

Orin "NX" is also an interesting name...

Nvidia has existing chips with "NX" in their name, and they're not the ones getting used for Switches. Dane will most likely be a variant that strips out a bunch of the extra accelerator hardware and swaps out A78AE for either A78 or A78C.

 

[Alovon11]

Good bye Lighting RT.

Hello audio RT.

RTXGI (NVIDIA's own RTGI Solution) can be scaled down to run on the Xbox One in software.

That can be accelerated by 4 Ampere RT Cores easily (Same with Lumen)

And there is the chance NVIDIA and Nintendo add back the missing 4 RT cores to make it a full 8 Again which would make it's RT performance near the Laptop 2060.

 

[lemonfresh]

It has the potential to be. The specs being discussed here are imo the best case scenario, it's likely to be less impressive than those. But still a good upgrade.

I'm curious how it will compare to something like the steam deck performance wise. if they release it like a year or two after I'm guessing it will be much better. But we will see

 

[Sol]

Nvidia has existing chips with "NX" in their name, and they're not the ones getting used for Switches. Dane will most likely be a variant that strips out a bunch of the extra accelerator hardware and swaps out A78AE for either A78 or A78C.

Still though.....8 A78C CPU cores, 12 GB LPDDR5, and a 1024 Cuda core Ampere GPU with 32 Tensor cores and 8 RT cores.....that would be an absolutely massive upgrade over the current Tegra Switch.

 

[Skittzo]

I'm curious how it will compare to something like the switch deck performance wise. if they release it like a year or two after I'm guessing it will be much better. But we will see

Based on what we're seeing here I'd say docked performance should be better. Maybe substantially so. Handheld mode is up in the air.

 

[YolkFolk]

I'm curious how it will compare to something like the steam deck performance wise. if they release it like a year or two after I'm guessing it will be much better. But we will see.

The system having DLSS should put it ahead of the Deck in terms of it’s overall output.

 

[Marce-chan]

The thing about branding that makes it hard for me to believe in a Super Nintendo Switch or Nintendo Switch Advance or something like that is because those heavily rely on nostalgia, and if they want to go the iterative route, NES-> SNES and GB -> GBA were rather super traditional next gens. Wouldn't be that ambiguous tbh.
Nintendo Switch S would sound like a cool branding, as S may sound like special, or super, etc. Also there's iPhone (insert number) S where the S implies it's a better model, and Galaxy S series being the top line. In video games tho there's Xbox One S, Series S, PS4(slim), so S may sound as a not-so-better revision.
Nintendo Switch Plus/+, or New Nintendo Switch, Nintendo Switch X(?), Idk, sound like something Nintendo would like to name it. Nintendo Switch S with a premium style would be neat too.

 

[ILikeFeet]

Still though.....8 A78C CPU cores, 12 GB LPDDR5, and a 1024 Cuda core Ampere GPU with 32 Tensor cores and 8 RT cores.....that would be an absolutely massive upgrade over the current Tegra Switch.

going by Orin, that would be 4 RT cores rather than 8. curiously, that means there are now 3 versions of Ampere

 

[Marce-chan]

Nintendo Switch without the heart

The Dark Souls of Nintendo Switch .

 

[Marce-chan]

I'm just curious: how do we know the chips that'll be used for Dane so early already?
Was there any leak specifically about Nintendo using the Orin chip, or part of it is speculation due to it coming late 2022/early 2023?

 

[lemonfresh]

I'm just curious: how do we know the chips that'll be used for Dane so early already?
Was there any leak specifically about Nintendo using the Orin chip, or part of it is speculation due to it coming late 2022/early 2023?

I'm guessing there are insiders among us

 

[Pokemaniac]

I'm just curious: how do we know the chips that'll be used for Dane so early already?
Was there any leak specifically about Nintendo using the Orin chip, or part of it is speculation due to it coming late 2022/early 2023?

A reliable Nvidia leaker said they're making a chip for Nintendo codenamed Dane that's in the general Orin family of chips.