NVIDIA Tegra Architecture

You'd hope the 16nm version at the very least. Was there much to speak of for Parker in terms of architecture (not counting the voodoo magic with the clocks)? TX1 already had double-rate FP16 for example. Pascal also updated the bandwidth compression I suppose.

 

Yeah unfortunately I don't think the game supports performing AF itself.

 

With a tablet one might assume they'd want the most efficient tech possible, though not necessarily a 256 ALU version of it like the known chips have. But it does appear to have active cooling (there is a vent) so I have a feeling it's at least 256 ALU. Tegra K1 has no problems gaming without active cooling.

 

If it's a K1 derivative then it's unlikely to have more than 192SPs. Both Tegra K1 and X1 have not problems gaming in ultra thin tablets for instance (without active cooling) as long as neither/nor GPU isn't clocked over 850MHz.
IMHO most likely candidate is some sort of X1 derivative under 20SoC TSMC.

Does anyone have a clue if there's another SoC in the dock?

 

It appears to have a fan. I saw at least one vent on the tablet in the video. That's potential for something more. Maybe just different clocks when docked though.

Or maybe the dock has the fan and moves air into it while docked...

 

Dosen't this imply that it's Pascal? Which would mean its 16nm. Wouldn't make sense to launch anything else in March 2017 IMHO.

Also do you really think Nintendo paid for a custom SoC or is this basically a Parker?

 

Don't they almost always go custom?

 

They do but what I meant was do you think its really a highly customised part or basically some minor tweak of Parker. Can't imagine they need too much more performance

 

Cannot tell from the info given so far but I would be expecting it to be a custom 16nm Pascal Tegra, after all we have the smaller Drive PX2 models now as well and logically they would also apply the experience and hardware to the Tegra world for certain products.
I am still curious about capabilities and optional hardware (if feasible) as they will be competing against next-gen mobile phones/tablets in early 2017, and they are more powerful than the Maxwell Tegra although those phones probably will catch fire or bend on you

Cheers

 

I suppose the Switch does change clock depending on whether it's on battery or on power socket, which would make sense given the change in resolution too...

 

I think it depends on whether they are using a secondary GPU in the station or not. You would need at least PCIE or something equivalent that is capable of unplugging during a running session. The Tegra itself has not very powerful I/O for external comms. I hope it is not only one Tegra, since it would be a very limiting factor for performance.

 

Tegra 3, K1, X1 and X2 all have PCI Express connections.
I don't think an external GPU in the dock makes a lot of sense because it would make it rather expensive and Nintendo officials have stated time and again that the new console would have to be affordable, but the limit would definitely not be from any I/O in current Tegra architectures.

 

PCIE 3.0 with at least 4 lanes?

 

All of them have 4 lanes. Parker / TX2 is most definitely PCI Express 3.0 since the Drive PX2 depends on external dGPUs.

 

Is there any technical paper to proof that? But yeah maybe you are right. Anyway: the story with the secondary GPU was denied by big N. So we wont see much better graphics...

 

https://gfxbench.com/result.jsp?ben...=true&arch-ARM=true&arch-x86=true&base=device

I suspect that the nVidia device at the top of the benchmark provides a decent guide for what to expect from the performance of the custom Tegra in the Nintendo Switch.

edit: Eh, looks like someone already posted it in this thread a while back. Never mind!

 

That's a Parker / Tegra X2.
Cudos to Nintendo and nvidia if that's the SoC driving the Switch.

It would stand at around half the performance of a PS4, or 2/3rds of a Xbone, so definitely enough for latest-gen visuals at 720p.
Given the expensive quad-channel LPDDR4 memory setup and 8GB of it nonetheless, plus the exquisite CPU core arrangement, I wouldn't count on it though.

 

They show the PCIex4 in their architecture slide, albeit I would say the GPUs seem to be dedicated per Parker.


Cheers

 

dGPU chips are significantly smaller than each Parker? LOL

 

I don't remember if the area for Parker has been disclosed.
In those diagrams, they have to put all the ports coming out of the SoC, and that's most probably why it's drawn larger than the GPUs (which are probably GP106)