Discussion Nintendo Switch and potential alternative ODM vendors

[soresu]

Its fine. Their relationship is good. They developed the NVN and NVN2 APIs together. Nintendo is the type of company that will cut ties immediately if their partners are not good. That fact they got the contract again for Switch 2 is impressive, Nvidia must have pitched hard.

I don't think it's that surprising or impressive at all.

It's sunk cost.

After so long with one rather similar hardware ecosystem before Switch, and having attracted so many new software partners with Switch they would need any alternative pitch to be insanely good to make it worth it at this point.

If Switch had tanked it's likely that they would have been more pliable to other offers.

What does surprise me is it sounds like Nintendo basically rushed Switch development.

I know that WiiU sales were less than exciting for them, but not even giving vendors time to produce a custom chip design for them seems to be a bit extreme.

 

[soresu]

I'm glad they used Maxwell, its what made that potato relevant even in 2024. Using Decaf would've been a nail in the coffin for Nintendo.

More like one more refrain in a long running joke.

(also the PowerPC 750 was far older than Maxwell in 2017 when Ninty finally shelved it)

One earlier refrain of the joke was the stunt they pulled during the WiiU hype cycle.

Implying that the WiiU CPU was the same POWER7 µArch used in IBM WATSON - a huge update from the PowerPC 750 used since Gamecube.

When the truth came out near release that it was just a souped up tri core PPC 750 I went from having planned to buy my first wall powered Nintendo console to giving up on them completely.

 

[poke01]

More like one more refrain in a long running joke.

the sun will hopefully set soon.

When the truth came out near release that it was just a souped up tri core PPC 750 I went from having planned to buy my first wall powered Nintendo console to giving up on them completely.

I only brought one cause of Splatoon and homebrewed it. Its a fun console but certainly died early in the market for a number of reasons.

Edited: so that it didn't read like my Wii U died

 

[soresu]

I only brought one cause of Splatoon and homebrewed it. Its a fun console but certainly died early for a number of reasons

I have considered getting Switch for the Link's Awakening remake, but it wouldn't be enough for that alone, even though I'd like to play a more recent Pokemon game too (haven't played a new one since the Gameboy Advance games).

If they made an OOT remake, and/or a MM remake and/or a Link to the Past remake in the style of Switch Link's Awakening I would definitely reconsider.

If they made all 3 I'd be

 

[marees]

I don't know if AMD could do an APU cheap enough for them while keeping power low.

No way it goes to AMD as TSMC has become very costly. Only Nvidia had/has a tie-up with Samsung

AMD are just dipping their feet in Samsung with Sonoma Valley chromebook APU

 

[Ghostsonplanets]

What does surprise me is it sounds like Nintendo basically rushed Switch development.

Yes


Contract for Tegra X1 and Nvidia being the SoC supplier was accorded around 2014 and console finals specs were finalized in Q2 2015 (ES). The only change the console specs went from was the Storage which was increased from 16GB to 32GB for the retail unit launch.

No way it goes to AMD as TSMC has become very costly. Only Nvidia had/has a tie-up with Samsung

We don't know the manufacturing foundry for T239 though. It might be fabbed on TSMC 4N or a newer Samsung node (unlikely imo).

T239, imo, seems a design that is too big for Samsung 8N. Post-Yamauchi Nintendo also never used such an old node for their new consoles launch.

 

[Ghostsonplanets]

BTW these were the specs for the Switch before Nvidia was the chosen SoC supplier:

4x Cortex A53
AMD Latte Decaf (AMD Latte 1/2)
1GB LPDDR3
Toronto SoC for 3DS BC
Miracast for Wireless Streaming from console to TV
853 x 480 Dual 3D Screens / Single Screen (with 120Hz option) (The screen spec kept changing throughout the development)

Target Power for SoC: 2W - 2.6W

I'm glad they used Maxwell, its what made that potato relevant even in 2024. Using Decaf would've been a nail in the coffin for Nintendo.

I agree. Using Decaf would have made for a more smoother progression from 3DS development teams and easy transpilation for Wii U titles. But for long-term business outlook, it would have been terrible.

 

[poke01]

fabbed on TSMC 4N

If this happens the Switch 2 might actually be day 1 one purchase, instead of lets wait for the holiday season.

 

[poke01]

https://twitter.com/x/status/1807726047468601459

Interesting, so that they would produce units at least 6 months early?

 

[Ghostsonplanets]

If this happens the Switch 2 might actually be day 1 one purchase, instead of lets wait for the holiday season.

Mind you, this is pure conjecture from me. But T239 seems exceedingly big for Samsung 8N and I don't see they being able to achieve Nintendo power targets (10W) on 8N.

Some people say Nintendo will just clock down o oblivion to make it fit on 8N. But then I question why make such a big design instead of making a smaller but faster clocked one which would yield higher and consume the same.

If it's not 8N, then that leaves only TSMC 7N or 4N as the node of choice because these are nodes Nvidia currently manufacture products on and have wafer procurement/supply.

Other route could be using a newer Samsung node (SF4+). But I find that unlikely as Nvidia wouldn't want to use a node that they have no other products being manufactured imo.

https://twitter.com/x/status/1807726047468601459

Interesting, so that they would produce units at least 6 months early?

Switch 2 is poised to sell 20+M units Y1. So, yes, that would make sense.

They also need to manufacture a lot of units to avoid Japan being scalped and leaving the country without consoles. That is because, as the yen keeps getting weaker against dollar, it means Nintendo will very likely have a lower price for Switch 2 in Japan compared to rest of the world.

 

[jpiniero]

Other route could be using a newer Samsung node (SF4+). But I find that unlikely as Nvidia wouldn't want to use a node that they have no other products being manufactured imo.

Switch is such a high volume product that I think that would be fine.

N4 would be too expensive. We are still talking about Nintendo here.

 

[Ghostsonplanets]

Switch is such a high volume product that I think that would be fine.

N4 would be too expensive. We are still talking about Nintendo here.

Yes. But if it happen that it's a failure, Nvidia is left sitting with a lot of unused wafer procurement. Like it happened with Renesas and Wii U.

N4 is expensive but is getting cheaper every year. For a product releasing in 2025, I wouldn't be surprised. I know it's Nintendo, but they also changed leadership twice since the Wii U. So I'm waiting to see how the new leadership acts.

 

[jpiniero]

There's always the possibility that future low end dGPUs are fabbed at Samsung.

Also I'd assume that unused wafers would be Nintendo's problem and not nVidia's.

 

[Ghostsonplanets]

There's always the possibility that future low end dGPUs are fabbed at Samsung.

Also I'd assume that unused wafers would be Nintendo's problem and not nVidia's.

Both greats points, yeah.

 

[soresu]

T239, imo, seems a design that is too big for Samsung 8N. Post-Yamauchi Nintendo also never used such an old node for their new consoles launch.

Great success with Switch allows them to be bolder in that regard.

They once took risks in the hardware realm as far back as Gamecube.

They certainly push the volume to make it worth it.

 

[ToTTenTranz]

T239, imo, seems a design that is too big for Samsung 8N.

T239 was designed in 2018-2020 for a hardware release in 2021, just like the big Orin chip T234 also on Samsung 8nm.

Other than some vague rumors that IMO sound more like wishful thinking than anything else, there's been no word of changes to the original 2021 design, much less redesigning the whole thing to now go into the high-end (and so much more expensive) TSMC N4.

As for size, the GA107 with 2560sp/20 SMP + 32 ROPs + 128bit GDDR6 + 8x PCIe is 200mm^2 big. The Switch 2's GPU (GA10F) is probably a bit over half the GA107 in size, with 1536sp/12SMP + 16ROPs. There's still 128bit LPDDR5 but then there's no need for 8x PCIe 4.0 in there and stuff like NVDEC/NVENC can probably be simplified as well.


My guess is the T239 is around 150mm^2, and that's fine for a tablet. What the 8nm means is the chip is going to be running really slow in its clock domains, and I wouldn't expect the GPU to go above 500Mhz in handheld mode, for example.

 

[Ghostsonplanets]

T239 was designed in 2018-2020 for a hardware release in 2021, just like the big Orin chip T234 also on Samsung 8nm.

2019. Tape-out in 2022.

Other than some vague rumors that IMO sound more like wishful thinking than anything else, there's been no word of changes to the original 2021 design, much less redesigning the whole thing to now go into the high-end (and so much more expensive) TSMC N4.

My thinking is that maybe it was designed for a more advanced node from start.

As for size, the GA107 with 2560sp/20 SMP + 32 ROPs + 128bit GDDR6 + 8x PCIe is 200mm^2 big. The Switch 2's GPU (GA10F) is probably a bit over half the GA107 in size, with 1536sp/12SMP + 16ROPs. There's still 128bit LPDDR5 but then there's no need for 8x PCIe 4.0 in there and stuff like NVDEC/NVENC can probably be simplified as well.

There's 4x PCIe 4.0 lanes (Nintendo uses them). Also the 8x Cortex A78 cluster. NVDEC and NVENC are full rate, same as Orin.

My guess is the T239 is around 150mm^2, and that's fine for a tablet. What the 8nm means is the chip is going to be running really slow in its clock domains, and I wouldn't expect the GPU to go above 500Mhz in handheld mode, for example.

Switch 2 is going to use a 7.91 1080p screen, so it will have a bigger body by default. That would support a bigger SoC fabbed on 8N. My question is more about power draw, battery life and also weight.

Nintendo will very much want to stay "Switch like" on these aspects. Which means a lower power draw, bigger battery life and reduced weight when compared to a standard PC Handheld. A more advanced node would help in these aspects.

 

[Ghostsonplanets]

Another thing to keep in mind is that T239 (Drake) wasn't being developed in parallel with Orin (T234). T239 project started after T234 and while it uses some of Orin blueprints, it also has extensive modifications. The memory controller for example is a LPDDR5X capable one for Nintendo SoC while Orin used a LPDDR5 memory controller IP.

 

[ToTTenTranz]

2019. Tape-out in 2022.

Original plan was most probably 2021.

https://twitter.com/x/status/1374260197724626944

Of course, these would be pre-COVID plans before the massive semiconductor shortages and massive Switch 1 sales that came from the lockdowns and social segregation.

My thinking is that maybe it was designed for a more advanced node from start.

Nah.

My question is more about power draw, battery life and also weight.

Rock bottom clocks, like the Switch.
If Orin's GPU clocks at 5-15W TDP are anything to go by, we're probably looking at ~400MHz handheld and ~800MHz docked.

The memory controller for example is a LPDDR5X capable one for Nintendo SoC while Orin used a LPDDR5 memory controller IP.

Changing one PHY for the other might be quite easy, especially glued to memory controllers that were originally designed to run GDDR6X at ~3x the clocks.
And I don't know if you can't just run LP5X on a LP5 PHY as long as they just limit the memory clocks to LP5 speeds and just enjoy the (lower) variable voltages.

 

[xpea]

Another thing to keep in mind is that T239 (Drake) wasn't being developed in parallel with Orin (T234). T239 project started after T234 and while it uses some of Orin blueprints, it also has extensive modifications. The memory controller for example is a LPDDR5X capable one for Nintendo SoC while Orin used a LPDDR5 memory controller IP.

Yep and the GPU is a also custom with some IPs from Ada for better RT/DLSS performance