Switch 2 Speculation

[Father_Murphy]

A couple of random thoughts.

First. The T239 was rumored/tipped as the SoC a ways back at this point. If it was as closely related to Orin as speculated, it was set for Samsung's 8nm process (same as Ampere presumably). Of course, it's entirely possible that's what we get. But, I wonder if Switch 2 was pushed back due to the Switch's success and, if so, if it became cost viable for T239 to be, at a minimum, ported to a better process or, even better, updated or replaced with a better CPU or with Ada cores. Too optimistic, I know.

Two. We know that Ampere has hardware that can enable frame generation, but the Nvidia line is that it was not high enough quality to enable on desktop. That's probably half true, and the other half is Nvidia wanted to reserve FG as a selling point for 4000 series cards. That incentive doesn't apply to the Switch 2. And, given the Switch 2 is a closed system, it may be possible and worth the investment for Nintendo / Nvidia to get FG working on Switch 2, at least for some titles.

 

[Chiimaero]

A couple of random thoughts.

First. The T239 was rumored/tipped as the SoC a ways back at this point. If it was as closely related to Orin as speculated, it was set for Samsung's 8nm process (same as Ampere presumably). Of course, it's entirely possible that's what we get. But, I wonder if Switch 2 was pushed back due to the Switch's success and, if so, if it became cost viable for T239 to be, at a minimum, ported to a better process or, even better, updated or replaced with a better CPU or with Ada cores. Too optimistic, I know.

Two. We know that Ampere has hardware that can enable frame generation, but the Nvidia line is that it was not high enough quality to enable on desktop. That's probably half true, and the other half is Nvidia wanted to reserve FG as a selling point for 4000 series cards. That incentive doesn't apply to the Switch 2. And, given the Switch 2 is a closed system, it may be possible and worth the investment for Nintendo / Nvidia to get FG working on Switch 2, at least for some titles.

T239 being 8nm is unlikely at this point. Say it ends up being TSMC 4nm, NVidia and Nintendo only would know the why. Switch success is an option. Not able to reach targets on 8nm is another. Beefed up targets (and thus clocks) is another. And so on. I'm sure we'll eventually get to know this "why" is due time (so, in ten years or so )
But a complete revamp with better parts is unlikely, someone would have find smth about a new chip at this point.

As for FG, I really don't think it can add value to the Switch 2 for technical reasons I discussed in my previous post. In short: FG is not well suited for power constrained devices, and brings too little to the table to begin with.

 

[Kaotik]

Yeah, it's confirmed it has the same DLSS hardware as Ampere cards (RTX30 series).

Let's be real, nothing has been confirmed. If it's the old Tegra there's no way it's hitting anywhere close to XB/PS levels

 

[Seanspeed]

T239 being 8nm is unlikely at this point. Say it ends up being TSMC 4nm, NVidia and Nintendo only would know the why. Switch success is an option. Not able to reach targets on 8nm is another. Beefed up targets (and thus clocks) is another. And so on. I'm sure we'll eventually get to know this "why" is due time (so, in ten years or so )
But a complete revamp with better parts is unlikely, someone would have find smth about a new chip at this point.

As for FG, I really don't think it can add value to the Switch 2 for technical reasons I discussed in my previous post. In short: FG is not well suited for power constrained devices, and brings too little to the table to begin with.

Switch 1 came out in 2017 on TSMC 20nm, which was a slightly older node nobody else really wanted to use and was hardly that great even for its time. But it was probably cheap, had the manufacturing capacity available, and was 'good enough'.

I dont think Samsung 8nm is out of the question for the same reasons. Especially if we start walking back any ideas that it's gonna be some super powerful platform, which I just dont see. That's just not what Nintendo does anymore and not only will they have a very 'affordable' price target in mind, but they also dont seem to like taking losses on hardware at any point like others will accept.

 

[Frenetic Pony]

Switch 1 came out in 2017 on TSMC 20nm, which was a slightly older node nobody else really wanted to use and was hardly that great even for its time. But it was probably cheap, had the manufacturing capacity available, and was 'good enough'.

I dont think Samsung 8nm is out of the question for the same reasons. Especially if we start walking back any ideas that it's gonna be some super powerful platform, which I just dont see. That's just not what Nintendo does anymore and not only will they have a very 'affordable' price target in mind, but they also dont seem to like taking losses on hardware at any point like others will accept.

It likely is far outdated, Switch didn't come out on TSMC20 because it was cheap, smaller nodes were still getting a bit cheaper back then still, and versus performance they still are getting cheaper today. Instead Switch came out with that because it was the failed Nvidia Shield SOC and Nvidia sold it cheap as dirt because they didn't know what else to do with it.

Samsung 4nm is cheaper than 8 for performance per $, and now Nvidia doesn't even have a real SOC line that can go into a console.

 

[Seanspeed]

It likely is far outdated, Switch didn't come out on TSMC20 because it was cheap, smaller nodes were still getting a bit cheaper back then still, and versus performance they still are getting cheaper today. Instead Switch came out with that because it was the failed Nvidia Shield SOC and Nvidia sold it cheap as dirt because they didn't know what else to do with it.

Samsung 4nm is cheaper than 8 for performance per $, and now Nvidia doesn't even have a real SOC line that can go into a console.

If Nintendo wanted to shell out the money, they could have asked Nvidia to port it to TSMC 16nm in the first place, as they eventually did later on. Point is - they were cheap about it.

As for Samsung 4nm being cheaper in performance per dollar, we have no idea what kind of offer Samsung might make to keep their 10/8nm manufacturing valid and profitable.

I get your point though. Situations aren't exactly the same. I'm not even necessarily saying I think it will be Samsung 8nm, just that it wouldn't shock me if it was. Nintendo are cheap bastards. lol

 

[T2098]

A couple of random thoughts.

First. The T239 was rumored/tipped as the SoC a ways back at this point. If it was as closely related to Orin as speculated, it was set for Samsung's 8nm process (same as Ampere presumably). Of course, it's entirely possible that's what we get. But, I wonder if Switch 2 was pushed back due to the Switch's success and, if so, if it became cost viable for T239 to be, at a minimum, ported to a better process or, even better, updated or replaced with a better CPU or with Ada cores. Too optimistic, I know.

Two. We know that Ampere has hardware that can enable frame generation, but the Nvidia line is that it was not high enough quality to enable on desktop. That's probably half true, and the other half is Nvidia wanted to reserve FG as a selling point for 4000 series cards. That incentive doesn't apply to the Switch 2. And, given the Switch 2 is a closed system, it may be possible and worth the investment for Nintendo / Nvidia to get FG working on Switch 2, at least for some titles.

There's more to the optical flow issue than just the quality of its output: https://docs.nvidia.com/video-techn...note/index.html#nvofa-quality-and-performance

Ada also significantly improved the speed of the optical flow analysis. If you look at the fps data on the right, Ampere, strangely enough, looks to be a regression in speed in many cases vs Turing, whereas Ada is 2x-3x faster. The footnote on the page also mentions that the 3 different GPUs were compared when locked at the same clock rate, so unless the OFA block runs on a different clock domain than the rest of the chip, it would lead me to believe Ada is even faster than indicated; it normally has a ~800mhz clock speed advantage over Ampere/Turing, give or take.

"All measurements are done by setting the video clocks as reported by nvidia-smi at 1755 MHz."

Combine that with the fact that even if the GPU in the Switch is Ampere architecture-wise, those mobile-style SoCs tend to run the GPU at extremely low clock rates for maximum power efficiency, kind of like the Max-Q laptop SKUs taken to their logical extreme.... that'd make the theoretical Ampere-class OFA block that was already too slow for frame generation (at least according to NV) when running at close to 2ghz on the desktop even less likely to be able to keep up at the closer to 1.0 to 1.2ghz clocks a mobile SoC like that would run at in the real world.

Who knows though, maybe they'll give it an Ampere GPU architecturally, but give it Ada's OFA block - they've played mix and match some of the secondary IP blocks in the past, like downgrading the NVENC/DEC block in TU117 to the one from Volta, versus giving it the full Turing one like all the other SKUs got.

 

[troyan]

It likely is far outdated, Switch didn't come out on TSMC20 because it was cheap, smaller nodes were still getting a bit cheaper back then still, and versus performance they still are getting cheaper today. Instead Switch came out with that because it was the failed Nvidia Shield SOC and Nvidia sold it cheap as dirt because they didn't know what else to do with it.

Samsung 4nm is cheaper than 8 for performance per $, and now Nvidia doesn't even have a real SOC line that can go into a console.

Nintendo used Tegra X1 because it was the most advanced graphics IP in the mobile space at the time. And it was more advanced than GCN 1.1 in the PS4 and Xbox. It was only cheap, because 20nm was the last process without FinFet.

 

[Frenetic Pony]

Nintendo used Tegra X1 because it was the most advanced graphics IP in the mobile space at the time. And it was more advanced than GCN 1.1 in the PS4 and Xbox. It was only cheap, because 20nm was the last process without FinFet.

Tegra X1 was 2 years old by the time Switch was released, it was ancient. Is your assumption really that companies only advance to newer nodes because they like paying more money than they need to?

 

[troyan]

And GCN was two years old, too, when Sony and Microsoft released their consoles. You should go back and read what was ready when Nintendo evaluated their supplier. Apple shipped the first 16nmFF SoC end of 2015, nVidia followed with Pascal in May 2016.

It was a Win2Win situation for both companies: nVidia had a ready to go SoC for mass production and Nintendo got access to the most advanced graphics IP which allows easy ports from PC and consoles.

 

[Kaotik]

And GCN was two years old, too, when Sony and Microsoft released their consoles. You should go back and read what was ready when Nintendo evaluated their supplier. Apple shipped the first 16nmFF SoC end of 2015, nVidia followed with Pascal in May 2016.

It was a Win2Win situation for both companies: nVidia had a ready to go SoC for mass production and Nintendo got access to the most advanced graphics IP which allows easy ports from PC and consoles.

They don't usually evaluate what's ready when they evaluate the suppliers, but what the suppliers will have ready when they plan to start mass producing the console.
GCN was two years old, but Sony and Microsoft didn't use GCN, they used GCN1.1 or "2nd gen GCN" which was fresh in 2013 when they launched. That's the normal way.

 

[Samwell]

Switch 1 came out in 2017 on TSMC 20nm, which was a slightly older node nobody else really wanted to use and was hardly that great even for its time. But it was probably cheap, had the manufacturing capacity available, and was 'good enough'.

I dont think Samsung 8nm is out of the question for the same reasons. Especially if we start walking back any ideas that it's gonna be some super powerful platform, which I just dont see. That's just not what Nintendo does anymore and not only will they have a very 'affordable' price target in mind, but they also dont seem to like taking losses on hardware at any point like others will accept.

There's a difference in using an already established SoC on an 3 year old node, which was done because nothing was changed from the already used Tegra, to a custom SoC. T239 might have many similarities to Orin, but it's still a completly different chip with a new tapeout and 8nm would be a 6 year old node. This is even much worse than Switch 1. Samsung 4nm might be too advanced, but 5nm would make pretty much sense. Samsung already lost Qualcomm and will try to get costumers with good prices and they will have the 4/5nm capacity as the production of a lot of stuff will move to 3 GAP next year.

 

[Silent_Buddha]

Some things for people to keep in mind when thinking of a new consoles to replace the NSW.

2 guiding principles at Nintendo that govern all hardware decisions.

• The games matter, the hardware only matters in as much as games need hardware to run on.
  • This doesn't mean the hardware is unimportant but the focus is and always will be on the games and not necessarily the hardware.
  • This also doesn't mean they won't point out cool things that their hardware can do or that they might not try to innovate with it (like with 3DS, for example with 3D display)
• The hardware should always be cheap enough that almost any family can afford one.
  • So, expect it to have a price around 349-399 at launch with a "lite" version either at launch or shortly after for 50 USD cheaper and base predictions upon that.
    • Nintendo also always sells their consoles with a nice profit margin unlike Sony and MS, so in Sony/MS terms this would be closer to a 249-299 hardware manufacturing and distribution costfor a Sony or MS console in terms of hardware costs.
      • NOTE - XBS-S was sold at a loss so it's hardware cost to manufacture and distribute was higher than 299 USD.
  • Tegra X1 for example was used because NV had a glut of inventory for it due to incredibly underwhelming demand in the markets it was created for.
    • Thus Nintendo was able to get it for far lower than NV would normally charge as NV wanted to clear inventory without just dumping it.
    • NV might also be kicking themselves right now for locking in a deal with Nintendo for the Tegra X1 with low margins, especially now that demand for their chips for automotive has exploded since then.
      • IE - it's highly unlikely that Nintendo will get another good deal for an NV based SOC.
  • So, a repeat of something like that might be unlikely.
    • IE - it's potentially possible that they go with AMD or possibly Samsung (AMD GPU IP available) or even another mobile chip maker.

In other words, the next Nintendo console won't be competing with PS5 or XBS-X and likely will be less powerful than the XBS-S. They'll be looking at hardware that can be manufactured and distributed for around 200-300 USD to be sold at retail for 300-400 USD.

Memory is expensive, I wouldn't expect a huge bump from 4 GB. 6 GB is likely and 8 GB might be possible, but I certainly wouldn't bet on anything higher than that.

Regards,
SB

 

[^M^]

You don't think the new console will be backward compatible with Switch game?

 

[babybumb]

The power narrative is getting way out of hand. Be suprised if it outperforms any on the current PC handhelds in any meaningful way.. other than battery life. Those run UE5 just "comparable" to PS5

 

[Kaotik]

You don't think the new console will be backward compatible with Switch game?

Changing architecture doesn't mean killing backwards compatibility

 

[Samwell]

• Tegra X1 for example was used because NV had a glut of inventory for it due to incredibly underwhelming demand in the markets it was created for.
  • Thus Nintendo was able to get it for far lower than NV would normally charge as NV wanted to clear inventory without just dumping it.
  • NV might also be kicking themselves right now for locking in a deal with Nintendo for the Tegra X1 with low margins, especially now that demand for their chips for automotive has exploded since then.
    • IE - it's highly unlikely that Nintendo will get another good deal for an NV based SOC.

Nvidia kicking themselves for the switch deal? Lower margins than they normally charge? The margins Nvidia had with the switch were great for a console deal and much better than AMDs. The question is, why Nintendo accepted such high margins for such an old SoC. Despite the investment in the auto segment and mediocre sales Nvidia made more profit with Tegra thanks to the switch than AMD with the Xbox One, One X, PS4 and PS4 Pro together. This is showing, which company was kicking themselves for the deal, but i can assure you it wasn't Nvidia. It's pretty easy to see, if you check their results and 10-K.

But Nvidia had the work upfront, because the API was coming from Nvidias Software department, unlike Xbox and PS. It seems Nintendo was desperate to get third parties onto the switch and therefore wanted Nvidias software help and accepted a very high price for the Tegra X1.

 

[Silent_Buddha]

Nvidia kicking themselves for the switch deal? Lower margins than they normally charge? The margins Nvidia had with the switch were great for a console deal and much better than AMDs.

Where'd you hear that? From buddy of mine that worked there at the time, word around the workplace was that NV was settling for at best single digit margins for the Tegra X1 provided to Nintendo. It was demoralizing for the team he was working on.

Regards,
SB

 

[Samwell]

Where'd you hear that? From buddy of mine that worked there at the time, word around the workplace was that NV was settling for at best single digit margins for the Tegra X1 provided to Nintendo. It was demoralizing for the team he was working on.

Regards,
SB

I haven't heard it, i'm reading their financial reports.
https://s201.q4cdn.com/141608511/fi...rterly_reports/2018/NVIDIA-Q4Y18-Form-10K.pdf

On Page 73 you have the profit of the tegra segment before and after the switch. Additional information are in the quarterly revenue trends, 10-Q reports from the quarters and 10K reports from further years and so on.
With the combination of the data you can see, that they made nice revenue and profit from it.

A possibilty is, they had two seperate deals. A low margin deal for the hardware and a seperate software deal, which made the deal profitable for nvidia.

 

[^M^]

Changing architecture doesn't mean killing backwards compatibility

Console API are rather close to the metal and changing architecture often means you need emulation which is not good for power consumption