Switch 2 Speculation

[Father_Murphy]

You definitely might be right, @Kise Ryota.

I tend to think Nintendo will position the Switch Pro/2 as a continuation of the Switch family, much like Nintendo successfully did with the Game Boy variants and then the DS family. In fact, Nintendo has arguably followed a version of the DS model with the Switch family. There was the DS (OG Switch) then a DS Lite (Switch Lite). Then, Nintendo released the DSi (upgraded internals), then the DSi XL (bigger screen). With the Switch, Nintendo flipped that, releasing the Switch OLED version before a version with upgraded hardware (e.g. a Switch Pro). With the DS family, games were generally compatible across all models with few exceptions (e.g. games requiring the DSi's camera which the DS/Lite lacked). Similarly, games are generally compatible within the Switch family (the exception being the Switch Lite's lack of removable Joycons).

The point is, if Nintendo continues a similar model, Nintendo would be launching the Pro in 2023 to preserve the momentum of the Switch. Nintendo would continue selling games to the massive Switch installed base (e.g. the new Zelda can be bought and played on any Switch). The idea would be to expand the base with a higher-end model, perhaps enticing families that already own a Switch to purchase another, or older gamers (like myself) to buy a new model simply because it is new and shiny !

 

[Kise Ryota]

Yeah, be it a pro or a switch 2, I also can totally see a hardware coming in 2023. Everything is possible.

But if it comes in 2023 1H, I'm not feeling it will be in 5nm (which kinda of brings me that same feeling of the switch in 20nm and not 16nm), which would be a shame.

For me, if they would release a Pro, it should have happened in 2020. But then shit happened. Now I think it's too late and they should just go with a proper next-gen. I also don't like the idea of a more smartphone like strategy (Drake is initially marketed as a Pro model, and 3 years after releasing it becomes the base model, and a new 'pro' is already coming, etc. I don't see Nintendo going for it.

But you see, sometimes it's just hard to have a clear picture when we have our own bias. I don't want Nintendo to go another route than the one I want for me, which can make harder to see the more logical path.

But Nintendo is unpredictable anyway lol
I think we need more substantial info going forward.

 

[Syferz]

This was just found, T239 has 8 cores in a single cluster, pointing to A78C being used and confirmed for new Switch hardware, this also might be used in a future Nvidia shield device, given the Linux OS.

So the device has 8 A78C cores and 1536 Cuda Cores and fits and runs in a Switch form factor... This really does seem like it won't be 8nm, if it is, I would expect the CPU to be underclocked a lot, still with 3 times the IPC vs A57, it will out do last gen consoles even with a very poor 1GHz clock, though I think a minimum clock for this would probably be more like 1.3GHz?

 

[orangpelupa]

You definitely might be right, @Kise Ryota.

I tend to think Nintendo will position the Switch Pro/2 as a continuation of the Switch family, much like Nintendo successfully did with the Game Boy variants and then the DS family. In fact, Nintendo has arguably followed a version of the DS model with the Switch family. There was the DS (OG Switch) then a DS Lite (Switch Lite). Then, Nintendo released the DSi (upgraded internals), then the DSi XL (bigger screen). With the Switch, Nintendo flipped that, releasing the Switch OLED version before a version with upgraded hardware (e.g. a Switch Pro). With the DS family, games were generally compatible across all models with few exceptions (e.g. games requiring the DSi's camera which the DS/Lite lacked). Similarly, games are generally compatible within the Switch family (the exception being the Switch Lite's lack of removable Joycons).

The point is, if Nintendo continues a similar model, Nintendo would be launching the Pro in 2023 to preserve the momentum of the Switch. Nintendo would continue selling games to the massive Switch installed base (e.g. the new Zelda can be bought and played on any Switch). The idea would be to expand the base with a higher-end model, perhaps enticing families that already own a Switch to purchase another, or older gamers (like myself) to buy a new model simply because it is new and shiny !

Switch oled, lite, and switch normal v2 are kinda like dsi. But nintendo deliberately didn't allow them to have higher performance.

 

[Pressure]

Jetson Orin Nano: 625MHz 512-core or 1024-core NVIDIA Ampere architecture GPU with 16 or 32 Tensor Cores and 6-core Arm Cortex-A78AE v8.2 64-bit CPU 1.5MB L2 + 4MB L3 in 5 to 15 Watt TDP.

That's the lowest end Orin. Then we can only be pleasantly surprised

 

[Dampf]

GPU wise, it sounds pretty awesome. But I am abit worried about the 8 core CPU, as it does not have any hyper threading. It could struggle in current gen exclusive third party games and CPU performance is much harder to scale than GPU. 12 cores would be better, to be on the safe side.

Or do these ARM cores have way higher IPC compared to Zen 2 in the consoles?

 

[Frenetic Pony]

Exciting news of the day:

GPU wise, it sounds pretty awesome. But I am abit worried about the 8 core CPU, as it does not have any hyper threading. It could struggle in current gen exclusive third party games and CPU performance is much harder to scale than GPU. 12 cores would be better, to be on the safe side.

Or do these ARM cores have way higher IPC compared to Zen 2 in the consoles?

It won't run the highest end CPU usage current gen games, at least not without a ton of optimization. AMD has better IPC than ARM and the CPU is likely to be 4big/4little if it's 8 core at all, so even less CPU power there. No building high end BVHs or huge physics sims or whatever.

Honestly don't care that much though. It's Nintendo, that's not what I buy Nintendo console for. If the Switch 2 can run Tears of the Kingdom at 1080p60 in mobile mode I'm more than happy.

 

[psurge]

Hmm - the rumored core count seems large for the target power envelope. Maybe they brought in a larger L2 from Ada to save on DRAM access power, or ported to another process? But for availability’s and all of our wallets’ sake, I hope it’s not 5nm - TSMC 6/7 would probably give a nice boost from Samsung 8 as well.

 

[Syferz]

Hmm - the rumored core count seems large for the target power envelope. Maybe they brought in a larger L2 from Ada to save on DRAM access power, or ported to another process? But for availability’s and all of our wallets’ sake, I hope it’s not 5nm - TSMC 6/7 would probably give a nice boost from Samsung 8 as well.

These aren't rumors, it's leaked. L2 and L3 is actually shared between CPU and GPU in Tegra, so the cache is huge regardless.

The SoC is known from Kopite7's tweet in spring 2021, where he identified T239 as a custom Tegra chip made for Nintendo. The Nvidia hack on March 1st this year had NVN2 in it (Nvidia's custom Nintendo API, as stated by them back in October 2016 with the Switch's reveal) and that code listed T239 as the only SoC that it uses.

In that code, it identified the GPU to have 1536 Cuda cores, 48 tensor cores and 12 RT cores, as well as 128bit bus.

Then today, T239 was spotted in a Linux update via open source database with a single 8 cpu core cluster, pointing to A78C as the only realistic option (the X series CPUs being the only other but they use too much power).

The information from today suggests T239 has been produced at least a month or two ago.

Lastly, I think this points to a smaller node than Samsung 8nm, whether that is Samsung 5nm, TSMC 6nm or Nvidia 4N, we don't know, but these 4 nodes are the options imo, and 8nm seems the least likely because you could make a smaller chip with a higher clock for the same power/performance ratio and save a lot of money.

 

[Goodtwin]

These aren't rumors, it's leaked. L2 and L3 is actually shared between CPU and GPU in Tegra, so the cache is huge regardless.

The SoC is known from Kopite7's tweet in spring 2021, where he identified T239 as a custom Tegra chip made for Nintendo. The Nvidia hack on March 1st this year had NVN2 in it (Nvidia's custom Nintendo API, as stated by them back in October 2016 with the Switch's reveal) and that code listed T239 as the only SoC that it uses.

In that code, it identified the GPU to have 1536 Cuda cores, 48 tensor cores and 12 RT cores, as well as 128bit bus.

Then today, T239 was spotted in a Linux update via open source database with a single 8 cpu core cluster, pointing to A78C as the only realistic option (the X series CPUs being the only other but they use too much power).

The information from today suggests T239 has been produced at least a month or two ago.

Lastly, I think this points to a smaller node than Samsung 8nm, whether that is Samsung 5nm, TSMC 6nm or Nvidia 4N, we don't know, but these 4 nodes are the options imo, and 8nm seems the least likely because you could make a smaller chip with a higher clock for the same power/performance ratio and save a lot of money.

Assuming all this hold true, a truly custom Orin chip should be able to free up a chunk of die space by removing all the automotive hardware. The AYA Neo units are shipping with AMD 6800U APU's, and they are 208mm2, double the size of the Tegra X1 in the Switch. I believe the original Orin chip die pic suggested it to be around 350mm2. Remove 4 CPU cores and 512 GPU cores along with the automotive hardware stuff and Nvidia should be able to ship a SOC about the same size as the AMD 6800U if they manufacture it at 6nm. Unlike the original Switch where Nvidia likely had logistical reasons for sticking with 20nm at launch before revising the X1 to 16nm less than two years later, with this new processor there is little reason to not choose a process gets the best performance to cost ratio that they can manage. Im really hoping this is coming in 2023, but if the specs are this good, it makes me feel like its going to be 2024.

 

[Father_Murphy]

That's the big question, isn't it. Is T239 a truly "custom" chip or is T239 "custom" in the sense that TX1 was said to be "custom" for the OG Switch but, in reality, amounted to a different binning of the chip.

 

[Syferz]

That's the big question, isn't it. Is T239 a truly "custom" chip or is T239 "custom" in the sense that TX1 was said to be "custom" for the OG Switch but, in reality, amounted to a different binning of the chip.

Nintendo was spotted working on TX1 power consumption in August 2014 via LinkedIn, so while TX1 was still not finalized, it was always a product for Nintendo, in fact there is an Nvidia Tegra roadmap from 2013 that says TX1 is codenamed "Parker" with Denver cores and Maxwell produced at 16nm.

A year later, there is a fork from Logan (K1) (Erista is Logan's kid) that Nintendo is working on.

 

[orangpelupa]

Ah so those code names does actually have real relevant meanings

 

[Syferz]

Exciting news of the day:
View attachment 6995

It won't run the highest end CPU usage current gen games, at least not without a ton of optimization. AMD has better IPC than ARM and the CPU is likely to be 4big/4little if it's 8 core at all, so even less CPU power there. No building high end BVHs or huge physics sims or whatever.

Honestly don't care that much though. It's Nintendo, that's not what I buy Nintendo console for. If the Switch 2 can run Tears of the Kingdom at 1080p60 in mobile mode I'm more than happy.

Just read through this post and thought I'd correct some things.

A78 has a higher IPC than Ryzen 2 found in Playstation 5 or Xbox Series. Also from the Linux code, we can see its a single 8 core cluster with a single frequency, if it were big.little or DynamIQ, there would be code for multiple frequencies for multiple types of cores. In fact there is even a comment about allocation of cache so that the kernal doesn't just load up 4 cores and properly distributes across all 8 evenly.

There is only 1 ARM chip that fits this bill and ARM created it for dedicated gaming hardware, that is the A78C version of ARM's Hercules cores.

As for BVH/Raytracing hardware, we know from the Nvidia hack that it has the far superior Ampere RT cores, while I don't expect much raytracing in handheld mode, because of DLSS, it should be capable of very good raytracing performance given its size.

 

[Pressure]

Just read through this post and thought I'd correct some things.

A78 has a higher IPC than Ryzen 2 found in Playstation 5 or Xbox Series. Also from the Linux code, we can see its a single 8 core cluster with a single frequency, if it were big.little or DynamIQ, there would be code for multiple frequencies for multiple types of cores. In fact there is even a comment about allocation of cache so that the kernal doesn't just load up 4 cores and properly distributes across all 8 evenly.

There is only 1 ARM chip that fits this bill and ARM created it for dedicated gaming hardware, that is the A78C version of ARM's Hercules cores.

As for BVH/Raytracing hardware, we know from the Nvidia hack that it has the far superior Ampere RT cores, while I don't expect much raytracing in handheld mode, because of DLSS, it should be capable of very good raytracing performance given its size.

Or the leak could be early internal hardware to test things out. There is just no point in speaking with confidence with something that is still probably a year or two out.

Just like Apple used A12Z to test out macOS working environment and application development before having M1 hardware on hand.

 

[Kaotik]

1536 Cuda cores scales fast. Steam Deck has 512 cores?

You can't just count cores of different architectures. If you look at desktops, we have AMD GPUs with 5120 "cores" competing with NV GPUs with 10240-10496 "cores"

 

[Syferz]

Or the leak could be early internal hardware to test things out. There is just no point in speaking with confidence with something that is still probably a year or two out.

Just like Apple used A12Z to test out macOS working environment and application development before having M1 hardware on hand.

So yesterday someone found in the Linux site that engineer samples of Drake existing in April this year. Rumors point to 1half of next year for the release, have for about 2 years now, both before and after bloomberg's report, and now Zelda is there in May.

You can't just count cores of different architectures. If you look at desktops, we have AMD GPUs with 5120 "cores" competing with NV GPUs with 10240-10496 "cores"

Yeah but that is sort of my point, we are talking about 3 times the cores and DLSS on top, digital foundry tried FSR2.0 on Drake and it wasn't fast enough too. This is just a lot of hardware, especially for Nintendo, so it is pretty exciting.

 

[function]

Going from past history, I'd expect Switch 2 to be around 12 ~ 18 months after general availability of the chip that's going to be used (or its contemporaries). So maybe late 2023 in time for a decent Christmas, or first half of 2024.

Particularly in mobile mode, clocks will be well below max clocks for the components. Nintendo will want reliable clocks, probably 3+ hours minimum battery life, and they won't want to spend a ton on the battery. At least that's my feeling anyhoo.

 

[psurge]

If it’s really a custom chip, does waiting that long make sense? I guess it gives devs time with actual HW, and maybe they are not in a rush with Switch selling so well. Anyway, if it’s an Ampere GPU, I doubt it will be 5nm - porting doesn’t sound trivial, and 5nm is expensive. Maybe TSMC 7, since A100 was produced on that process. (But my bet is they stay with Samsung 8.)

 

[Frenetic Pony]

Going from past history, I'd expect Switch 2 to be around 12 ~ 18 months after general availability of the chip that's going to be used (or its contemporaries). So maybe late 2023 in time for a decent Christmas, or first half of 2024.

Particularly in mobile mode, clocks will be well below max clocks for the components. Nintendo will want reliable clocks, probably 3+ hours minimum battery life, and they won't want to spend a ton on the battery. At least that's my feeling anyhoo.

Batteries are the heaviest component of any mobile tech these days, it's over half a phone's weight. And the Switch's battery life notably kinda sucks, there's a very large percentage of customers that would really prefer another hour of battery life over "moar pixels" if there's a tradeoff, and extra battery adding cost and weight doesn't recommend itself to being made by Nintendo.

More interesting is the controller. The Switch's is meh rated. Needs to be more comfortable, less prone to accidentally clicking the sticks (first controller I've ever had that problem) drift, etc. Combined with the PS5 Dualsense being pretty cool, and the possibility of making the whole thing a VR headset with an addon (120hz 1080p VRR OLED screen) and... well there's way more relevant improvements to be made in the controller area. VR Metroid Prime 4, VR New Pokemon Snap, and VR Skyward Sword already suggest themselves pretty highly.