Switch 2 Speculation

Whatever they do , I hope the new switch can play the older games with significantly longer battery life.

I have a switch but could get a switch 2 if its powerful enough. Unless one of the other players come out with something better

IIRC historically Nintendo did weird things like Nintendo DS low battery life, upgraded on DS lite. Ut downgraded again on DSi.

As for switch, it got upgraded in switch 1v2.
 
As for GPU, that's a lot harder. The iphone 11, the current champ, only delivers about 330 max fps on T-rex offscreen.
And throttles down to 270 after a few minutes. Heat dissipation without active cooling is a limiting factor, and you can't put a fan in there as it would drain too much power.
While the roughly equivalent to a PS4 7850 hits 410 fps. However, hypothetically Imagination's upcoming new GPU arch can match the PS4. And if this benchmark is to be believed then AMD, and one would assume Nvidia, could match it at 181fps in Manhattan, 20 above a 7850, and the Switch 2 would have a higher TDP to do it in.
One can also factor in efficiencies. You could use less TF of RDNA to achieve the same on-screen results as 1.4 TFs of PS4. Something comparable to PS4 in mobile form 8 years after its release is certainly realistic. How close it can get to next gen (things like fast RAM or storage) is going to be restricted by battery and thermals. eg. Could it really use super-fast 3+ GB/s SD cards, or would they just get too hot when there's zero airflow across them?
 
IIRC historically Nintendo did weird things like Nintendo DS low battery life, upgraded on DS lite. Ut downgraded again on DSi.

As for switch, it got upgraded in switch 1v2.
I remember the wii u pad having a tiny battery but space for a much larger one. So who knows.
 
Here’s my Switch 2 speculation that was inspired in part by the DF DLSS videos.

I think instead of running the GPU at a higher clock rate in docked mode, simply run DLSS on the tensor cores. DLSS is probably even more suitable for Nintendo style graphics.

So my prediction is a NVIDIA chip with 2x the perf of the current Switch and enough tensor cores to upscale 720/60 to 4K/60.
 
I don't see any portable device having Tensor cores at all. Right now it doesn't seem to fit the power envelope.

So what can Nvidia do to meet those strict requirements?
 
Instead of dlss, Nintendo could use amd fidelityfx as it doesn't need tensor core and the result was good enough
 
I don't see any portable device having Tensor cores at all. Right now it doesn't seem to fit the power envelope.

So what can Nvidia do to meet those strict requirements?

But (honest question) what cost most power and space, more rop / textures units / shaders, or some tensor cores ?
 
Any reason not to build it into the dock?

As in a image post-processor or as in an external GPU that is used instead of the integrated GPU? Or some sort of external device the game feeds the needed inputs into the DLSS ML Algorithm, like the old school FPU Coprocessors?

I'm not certain, but I was under the impression that DLSS requires more than just the screen space representation of the image. @iroboto any thoughts on this fun little side speculation?
 
But (honest question) what cost most power and space, more rop / textures units / shaders, or some tensor cores ?

Really not sure on this, at least not what will be in the Nvidia 30x0 series of GPUs, or even the power gating aspects of it. Maybe they have ability to completely turn that silicon dark and not consume any power whatsoever when not used. Also not certain on what the minimum number of tensor cores required to make it usable.

It certainly is fun to speculation on what they may cook up.
 
I don't see any portable device having Tensor cores at all. Right now it doesn't seem to fit the power envelope.

So what can Nvidia do to meet those strict requirements?

Every single mobile SOC already has (or will in the next iteration) ML accelerators for inferencing. Providing a dedicated ML inferencing engine is much more efficient in terms of power, area, and performance than doing that on the CPU or GPU.

ML upscaling is a very hot topic among mobile vendors since it offers a tremendous power savings compared to having to scale the GPU up for rendering. With in a couple of years, I bet you will see DLSS like solutions across the board in the mobile space running on dedicated hardware (call it ML engine, tensor core, etc).
 
Here’s my Switch 2 speculation that was inspired in part by the DF DLSS videos.

I think instead of running the GPU at a higher clock rate in docked mode, simply run DLSS on the tensor cores. DLSS is probably even more suitable for Nintendo style graphics.

So my prediction is a NVIDIA chip with 2x the perf of the current Switch and enough tensor cores to upscale 720/60 to 4K/60.

That is not that bad of an idea. Add custom processing on the doc to simply upscale the frame intelligently to 4K. That can be DLSS, or some other thing, and it can be done through Tensors, regular ALU/shader units, or some custom build HW for that precise purpose.

Also, nothing keeps nintendo fron building a Switch 2 IO interface that lets it send more than just a framebuffer. If the upscalling algo would benefit from motion vectors, stencil masks, amd maybe UI and HUD in a separate layer, then just build that into their system. Why not?
 
As in a image post-processor or as in an external GPU that is used instead of the integrated GPU? Or some sort of external device the game feeds the needed inputs into the DLSS ML Algorithm, like the old school FPU Coprocessors?

I'm not certain, but I was under the impression that DLSS requires more than just the screen space representation of the image. @iroboto any thoughts on this fun little side speculation?

I was under the impression that it could use the screen space alone. That is probably incorrect now that I think about it. I was thinking along the lines of an external device. The internal GPU would still perform all the usual tasks, just the output would be run through Tensor-like cores for ML upscaling in the dock before heading to the screen.

And just now I saw McHuj's post about mobile being heavily interested. Maybe it could simply be built in.
 
That is not that bad of an idea. Add custom processing on the doc to simply upscale the frame intelligently to 4K. That can be DLSS, or some other thing, and it can be done through Tensors, regular ALU/shader units, or some custom build HW for that precise purpose.

Also, nothing keeps nintendo fron building a Switch 2 IO interface that lets it send more than just a framebuffer. If the upscalling algo would benefit from motion vectors, stencil masks, amd maybe UI and HUD in a separate layer, then just build that into their system. Why not?

I doubt Nintendo is going to put any hardware in the dock. Much simpler to put it on the SoC. Hardware in the dock is just going to make it more complex and more expensive because now you have multiple chips, you need a interface between the Switch and dock with enough bandwidth etc. If you put the tensor cores or whatever on the SoC you can also use them in mobile mode which sounds very handy to me given the power constraints

I also rather see Nintendo aim for good 1080p than half assed 4k. I tried some games on my 55" tv and I can't say I can see a very obvious difference between 1080p and 4k.
 
Anyone knows how Wii U handled the GPU under Wii mode? Because Maxwell's ISA can be quite different than Volta's and onward. If games on Switch are shipped with precompile shader then with new GPU they will have a compatbility problem to solve.
 
Anyone knows how Wii U handled the GPU under Wii mode? Because Maxwell's ISA can be quite different than Volta's and onward. If games on Switch are shipped with precompile shader then with new GPU they will have a compatbility problem to solve.

AMD included the Flipper/Hollywood GPU on the die along Latte GPU so that's how backwards compatibility was achieved ...

If Nintendo's new system doesn't include the Maxwell/Pascal ISA then RIP backwards compatibility unless they integrate the Tegra X1 GPU either on die or off die which would involve a lot of redundant functionality taking up more die space/chipset size ...

Volta+ ISA is pretty much incompatible with the previous generation's ISA. If Nintendo wants some sort of forward compatibility then they should partner with AMD on GPUs but that strategy won't start paying off until they release a new successor again ...
 
Good emulation can be achieve I guess ? Anyway, I don't know if Nintendo customers care about bc when the hardware is so different from a generation to another (wiimote, then the screen/pad, then the switch...)
 
Good emulation can be achieve I guess ? Anyway, I don't know if Nintendo customers care about bc when the hardware is so different from a generation to another (wiimote, then the screen/pad, then the switch...)
That would mean dynamically translating Maxwell instructions to post-Volta instructions, which will be computationally expensive. If it’s going to be incompatible then they shouldn’t brand it as Switch 2/Pro...
 
Ah, I guess RIP backward compatibility then....

Or, or.........the very Nintendo thing to do is go with a custom processor that is a direct evolution of the Tegra X1 and thus no issues with backwards compatibility. Increase the memory bus to 128 bit matched up to the best LPDDR memory available, increase the shader ALU's to 512, clock speeds up to 1500Mhz thanks to a much smaller manufacturing process, and run those A57 CPU cores at 2Ghz. Incorporate the AA processor Mcable uses into the dock, and bam, you have Switch 2. This should deliver PS4 performance in the form factor of the Switch, and avoid compatibility issues. The Tegra X1 was one of the most modern processors Nintendo has used since the Gamecube. Thinking that Nintendo will chase peak performance just isnt something they do.
 
Back
Top