Based on how tech has progressed since the Switch, how powerful could a Switch Pro in early 2021 actually be?

[Skittzo]

Is TX1 gone for good according to recent developments? It wouldn't be a new 3DS situation if that was the case.

Click to shrink...

The recent info only really indicated that Mariko is getting discontinued sometime in the near future. Whether that means TX1 is gone or whatever new chip coming is another shrink/variant of TX1 is currently a matter of speculation.

Nate mentionned he was told it was going to be on A78

Click to shrink...

Right, that was one of the things he was told that he couldn't vet.

 

[Jraphics Horse]

Can someone explain the TSMC situation? The reports I saw mentioned they are ending a 3% discount, but is that it? On its own it does not seem like something that would cause a huge shift in plans.

 

[ILikeFeet]

Has any rumor actually mentioned A78 or is that just where speculation is because it makes sense? A few pages ago someone posted a rumor back from August I think which indicated Samsung presented Nintendo with A76s but not A78s.

Click to shrink...

As I have been corrected, Samsung has like 3 cpus on 10/8nm. The oldest is at least 2x more powerful than the A57

 

[Skittzo]

Can someone explain the TSMC situation? The reports I saw mentioned they are ending a 3% discount, but is that it? On its own it does not seem like something that would cause a huge shift in plans.

Click to shrink...

Someone who is apparently in the know about these things has been told that TSMC's 12nm production lines will be ending in early January. I don't remember if that post came before or after the news about TSMC ending their discount but it lines up with that info.

Whether or not you want to believe that person is up to you but that's what all of the current speculation is based on.

 

[ILikeFeet]

Can someone explain the TSMC situation? The reports I saw mentioned they are ending a 3% discount, but is that it? On its own it does not seem like something that would cause a huge shift in plans.

Click to shrink...

Due to high demand, they're gonna end their high volume discount and maybe increase prices. That 3% is probably a per wafer discount, not a total volume discount, so that adds up when you need several million chips

 

[Jraphics Horse]

Someone who is apparently in the know about these things has been told that TSMC's 12nm production lines will be ending in early January. I don't remember if that post came before or after the news about TSMC ending their discount but it lines up with that info.

Whether or not you want to believe that person is up to you but that's what all of the current speculation is based on.

Click to shrink...

I don't see a reason to connect the two things unless this was known years in advance, but that seems unrealistic.

No idea about the info, they seem to have picked it up as one of many rumours on a 'French community' whatever that means. If we assume the usual of it being a Ubisoft leak, I'd wonder if it relates to the end of the recent Mariko devkits rather than TX1 itself.

 

[Thugstas]

Witho

Click to shrink...

Nate mentionned he was told it was going to be on A78

Click to shrink...

True 😉

 

[Skittzo]

I don't see a reason to connect the two things unless this was known years in advance, but that seems unrealistic.

No idea about the info, they seem to have picked it up as one of many rumours on a 'French community' whatever that means. If we assume the usual of it being a Ubisoft leak, I'd wonder if it relates to the end of the recent Mariko devkits rather than TX1 itself.

Click to shrink...

The reason to connect it is that Nintendo is likely TSMC's only customer for 12nm chips at the moment. So them raising prices next year could very reasonably lead to Nintendo dropping them, and therefore dropping production of 12nm chips at the same time.

And I'm sure all of that has been set in stone for a while.

 

[takriel]

So has Thugstas contradicted themselves already?

 

[ILikeFeet]

The reason to connect it is that Nintendo is likely TSMC's only customer for 12nm chips at the moment. So them raising prices next year could very reasonably lead to Nintendo dropping them, and therefore dropping production of 12nm chips at the same time.

And I'm sure all of that has been set in stone for a while.

Click to shrink...

they aren't. I recall seeing some other products on TSMC's 12nm, so it's not exclusive to Nvidia

 

[Skittzo]

So has Thugstas contradicted themselves already?

Click to shrink...

Not that I've seen...

they aren't. I recall seeing some other products on TSMC's 12nm, so it's not exclusive to Nvidia

Click to shrink...

Ah is that so? It'd be interesting to see who else uses it, that would make it easier to see if there's truth to that claim.

 

[ILikeFeet]

Ah is that so? It'd be interesting to see who else uses it, that would make it easier to see if there's truth to that claim.

Click to shrink...

Mediatek has a line of chips that use 12nm

Fabrication

www.mediatek.com

Helio P30 | Extensive Dual-Camera System

MediaTek Helio P30

www.mediatek.com

 

[Lwill]

So has Thugstas contradicted themselves already?

Click to shrink...

With what? From a quick search of the poster's history, he implied that Samsung was involved in this back in September. I'm unsure if he is implying that Nate is correct with the A78, but the specific chipset wasn't mentioned from him in the previous posts afaik.

 

[Thugstas]

So has Thugstas contradicted themselves already?

Click to shrink...

🤔

 

[Thugstas]

With what? From a quick search of the poster's history, he implied that Samsung was involved in this back in September. I'm unsure if he is implying that Nate is correct with the A78, but the specific chipset wasn't mentioned from him in the previous posts afaik.

Click to shrink...

Knew bout A78 process months ago from arm.

 

[Gurgelhals]

The reason to connect it is that Nintendo is likely TSMC's only customer for 12nm chips at the moment. So them raising prices next year could very reasonably lead to Nintendo dropping them, and therefore dropping production of 12nm chips at the same time.

And I'm sure all of that has been set in stone for a while.

Click to shrink...

I find that hard to believe to be honest. According to TSMC's recent financial statements, their 16nm node (which includes the 12nm node, as this one is simply an optimised version of the 16nm process) still amounts to 18% of their total revenue. In all likelihood, they still have a lot of customers depending on that node and they can't just shut it down overnight. Also, it's something that TSMC doesn't really do in general. They still have fabs churning out 28nm and 40/45nm chips as well and these processes are ancient, comparatively speaking – while still contributing about 20% to their total revenue.

Maybe someone has the 16/12nm node confused with their half-step 20nm or 10nm nodes. Barely anyone seems to be using those, so maybe TSMC will shut those down.

 

[Thugstas]

Knew bout A78 process months ago from arm, no mention of it was for Ninty 😉

Click to shrink...

 

[Look over there]

There's also the question of, who does the A78C exist for. The major change is being able to stick up to 8 big cores within 1 cluster. As far as I can think of, the main workload on a mobile chip that'd benefit from that is gaming. To quote BlackTangMaster again:
"I have the feeling that the A78C have been designed specifically for gaming in the first place. ARM customers have all the data concerning newer cores long before the announcement of said cores. Nvidia is working on A78 since Xavier release.
The announcement of A78C is probably one of the most important information we've got concerning the new model planned for 2021. Outside of Nvidia (for a nintendo switch), I don't see who would want to make a switch like device, rather it is a full dedicated gaming device or a gaming smartphone. Either nintendo will have competition starting this year. Or these cores were made for the Nintendo switch.
In any case, the only ones who know how to make SOCs with more than 4 A78s are already those who produce chips for Nintendo. And I expect any laptop manufacturers to use one cortex X1 due to its single threaded performance. Something the A76, and thus the A78, will probably lack as seen on Qualcomm chips made for Windows on ARM."

---

Ok, trying to wrap my head around the flops conversion from Maxwell w/ mixed precision, without mixed precision, and Ampere.
First off, to get the target 1.5 tflops with Ampere, it's about 6 SMs at ~1ghz each. So triple the SMs and up the clock by a 1/3, or a combined multiplier of 4x.
This is roughly equivalent to ~900 X1 gflops with mixed precision taken advantage of. That is, take the Switch, then through a mix changing SM count and clock rate, hit a multiplier of ~2.25.
For games that don't utilize mixed precision, it's equivalent to ~1350 Maxwell gflops. ~1350/396 ~= 3.4. Something's off here, because perf per SM should not be going downward with a newer architecture. I feel like my math derailed at some point; probably back with the Ampere calculation?

 

[Skittzo]

I find that hard to believe to be honest. According to TSMC's recent financial statements, their 16nm node (which includes the 12nm node, as this one is simply an optimised version of the 16nm process) still amounts to 18% of their total revenue. In all likelihood, they still have a lot of customers depending on that node and they can't just shut it down overnight. Also, it's something that TSMC doesn't really do in general. They still have fabs churning out 28nm and 40/45nm chips as well and these processes are ancient, comparatively speaking – while still contributing about 20% to their total revenue.

Maybe someone has the 16/12nm node confused with their half-step 20nm or 10nm nodes. Barely anyone seems to be using those, so maybe TSMC will shut those down.

Click to shrink...

Okay yeah if 16nm/12nm still has a ton of customers then I doubt they're ending all of it.

 

[ILikeFeet]

The A78 is also good for laptops as well, but ARM has been putting "mobile gaming" before laptops in their promo material. There ain't any other gaming devices unless Ouya is coming back. Maybe a new ROG phone?

 

[z0m3le]

There's also the question of, who does the A78C exist for. The major change is being able to stick up to 8 big cores within 1 cluster. As far as I can think of, the main workload on a mobile chip that'd benefit from that is gaming. To quote BlackTangMaster again:
"I have the feeling that the A78C have been designed specifically for gaming in the first place. ARM customers have all the data concerning newer cores long before the announcement of said cores. Nvidia is working on A78 since Xavier release.
The announcement of A78C is probably one of the most important information we've got concerning the new model planned for 2021. Outside of Nvidia (for a nintendo switch), I don't see who would want to make a switch like device, rather it is a full dedicated gaming device or a gaming smartphone. Either nintendo will have competition starting this year. Or these cores were made for the Nintendo switch.
In any case, the only ones who know how to make SOCs with more than 4 A78s are already those who produce chips for Nintendo. And I expect any laptop manufacturers to use one cortex X1 due to its single threaded performance. Something the A76, and thus the A78, will probably lack as seen on Qualcomm chips made for Windows on ARM."

---

Ok, trying to wrap my head around the flops conversion from Maxwell w/ mixed precision, without mixed precision, and Ampere.
First off, to get the target 1.5 tflops with Ampere, it's about 6 SMs at ~1ghz each. So triple the SMs and up the clock by a 1/3, or a combined multiplier of 4x.
This is roughly equivalent to ~900 X1 gflops with mixed precision taken advantage of. That is, take the Switch, then through a mix changing SM count and clock rate, hit a multiplier of ~2.25.
For games that don't utilize mixed precision, it's equivalent to ~1350 Maxwell gflops. ~1350/396 ~= 3.4. Something's off here, because perf per SM should not be going downward with a newer architecture. I feel like my math derailed at some point; probably back with the Ampere calculation?

Click to shrink...

A78C is a design that doesn't lend itself to anything but dedicated Gaming consoles imo, and when you read about the CPU it specifically says it is designed for gaming. ARM is also no stranger to making customized CPU cores for Nintendo, which was done for Game Boy "Atlantis" and more or less became GBA. A78C is also a pretty small customization.

As for Ampere Architecture, I explained it, but compare any RTX 30 series card to the RTX 20 series, the RTX 3070 is a 20TFLOPs GPU but only matches the RTX 2080ti which is a 14TFLOPs Turing GPU, I explained it, Ampere's theoretical performance is bloated because the higher number comes from doubling the number of Cuda cores, nothing else in the SM saw nearly as much of a performance increase, so the Cuda cores are largely idle, however because those extra Cuda cores comes from the int32 operators which are usually idle in Turing anyways, so really if it needs to be boiled down this way, you could look at Ampere's GPU when compared to Tegra X1, as a 768GFLOPs GPU with mixed precision added automatically without needing to code for it... Of course it's not truly that way, theoretically, it is indeed 1.536TFLOPs but that would be the difference.

Also most Maxwell designs are 64 Cuda cores per SM, Tegra X1 is a very unique design and had 128 Cuda cores, if you compare it to tx2, they had the same GPU performance per clock but tx2 is a 4 SM GPU.

 

[NineTailSage]

Ampere has better power envelope than Radeon Vega, yet the 4500u is a 15 watt APU that has 6 cores at 2GHz + 6 CU (same as SM) at 1.5GHz. So no, Ampere just has inflated numbers, my entire post is about how 1.5TFLOPs Ampere is about the same performance as 900GFLOPs Tegra X1 (when mixed precision is involved).

This is the Tensor cores, not Cuda Cores, since at least Turing, Nvidia dropped what AMD calls rapid packet math, or doing 2 fp16 math ops for every fp32 ALU per clock.
You can see it here in TechPowerUp's run down of the specs under "Theoretical Performance" box, the half precision (fp16) and full precision (fp32) is both 29.77TFLOPs on the RTX 3080, but it's the same case with all RTX cards, and even the GTX 1600 series iirc.

Yeah, I at least know that testing was done, and the target was 2.25x base docked clock, but that performance included mixed precision, which is overlooked.

Click to shrink...

I definitely understand the Ampere having a better power envelope, but the Radeon Vega is also on TSMC's 7nm.
We see how the RTX 30 series cards perform (TDP wise) and there's no issue with the graphics architecture, the 8nm process is just showing it's age even in comparison to Samsung's own 7nm process (which is theoretically 40-50% more power efficient than 8nm).

In a prior podcast I've touched on the idea of a Lite connecting to a TV via a wireless docking solution similar to AirPlay.

Click to shrink...

I still think this is the much needed docking evolution for Switch...
It's seamless and definitely looks more high tech than manually docking the system.

I'm still on team 4-core, but an 8-core is also on the table. hell, a 6-core is also on the table. if I remember correctly, the PS4 had 5 cores available to devs so a 6-core wouldn't be that different

Click to shrink...

When you say 4 cores are you talking about 4-A78 cores + 4-A55? Or are you just discussing the amount of A78 based for the CPU. A 4-A78 + A55 set-up would be a major upgrade over both the PS4/XboxOne.

I find that hard to believe to be honest. According to TSMC's recent financial statements, their 16nm node (which includes the 12nm node, as this one is simply an optimised version of the 16nm process) still amounts to 18% of their total revenue. In all likelihood, they still have a lot of customers depending on that node and they can't just shut it down overnight. Also, it's something that TSMC doesn't really do in general. They still have fabs churning out 28nm and 40/45nm chips as well and these processes are ancient, comparatively speaking – while still contributing about 20% to their total revenue.

Maybe someone has the 16/12nm node confused with their half-step 20nm or 10nm nodes. Barely anyone seems to be using those, so maybe TSMC will shut those down.

Click to shrink...

Someone else brought this up that other manufacturers out there still have the 28nm 16nm and 12nm process going, which means TSMC has more competition on those nodes. They could be migrating production away from these older processes to make more room to upgrade more fabrications to 7nm, 5nm and whatever is next for TSMC.

TSMC knows that they can make way more money on the higher end smaller nodes and that they have less competition while doing it...

 

[UltraMagnus]

1.5 TFLOPS portable with 50-68GB memory bandwidth that probably has to cost $350-$400 I think will just run into the same "problem" that it's supposed to be solving ... in 2 years (probably even less) many devs will say that's not good enough for modern 3rd party ports as they move into PS5/XSX specific development.

It also needs to run in the 11-12 watt range too, maybe 15 watts max like the current Switch does, things like 17-18 watts just for the GPU is way too high.

 

[ILikeFeet]

When you say 4 cores are you talking about 4-A78 cores + 4-A55? Or are you just discussing the amount of A78 based for the CPU. A 4-A78 + A55 set-up would be a major upgrade over both the PS4/XboxOne.

Click to shrink...

Just 4 A78 cores in a not exactly worse case scenario. I still don't see A55s being used

 

[Gurgelhals]

Someone else brought this up that other manufacturers out there still have the 28nm 16nm and 12nm process going, which means TSMC has more competition on those nodes. They could be migrating production away from these older processes to make more room to upgrade more fabrications to 7nm, 5nm and whatever is next for TSMC.

TSMC knows that they can make way more money on the higher end smaller nodes and that they have less competition while doing it...

Click to shrink...

The 28nm, 16nm, etc. nodes of TSMC's competitors (mostly Samsung and GlobalFoundries) aren't compatible though. Switching from one manufacturer to another or even from one process to another usually requires a complete respin of the chip itself. Basically, once you've taped out a chip design with a particular manufacturer on a particular node of theirs, you're on that node and with that manufacturer unless you decide to "port" your design to another process and/or manufacturer. And this is not a simple process, but quite expensive and takes a least a few months. So TSMC shutting down what is still one of their biggest nodes (revenue-wise) would just leave a lot of their customers stranded (and pissed off) and it's not really good business sense for that reason alone. There are a lot of non-high-end chipsets being produced on those older nodes which work just fine and where there's absolutely no need to shrink them to a smaller node or anything.

Moreover, phasing out older nodes doesn't really mean more room for their 7nm and 5nm nodes. As far as I know, it's not really possible and/or feasible to upgrade older production lines to newer processes. To the best of my knowledge, these fabrication units have to be built pretty much from scratch every time and that's a multi-year endeavor anyway (just take a look at how Intel's manufacturing is pretty much fucked for at least a few years because their 10nm process didn't live up to their own expectations). So shutting down older processes would mean absolutely nothing wrt their capacity to deliver 7nm and 5nm chipsets in the short term. Yes, demand for various 7nm products far exceeds what TSMC can currently churn out, but there's really no short-term solution here. It really boils down to high-end semiconductor manufacturing being a seller's market at the moment and not a buyer's market.

 

[Deleted member 37917]

That's what the Switch is regardless of its power threshold. Its literally built with a portable-first design.

Click to shrink...

As a pure portable, the switch is kind of meh, in my opinion. Uncomfortable controllers, poor battery life (though better with the revision), bulky...
As a home console, it's kind of underwhelming too of course.

But both together, it was pretty great when it first released. A true hybrid. Next year's model should help recapturing that.

 

[John Omaha]

1.5 TFLOPS portable with 50-68GB memory bandwidth that probably has to cost $350-$400 I think will just run into the same "problem" that it's supposed to be solving ... in 2 years (probably even less) many devs will say that's not good enough for modern 3rd party ports as they move into PS5/XSX specific development.

Click to shrink...

In 3-4 years they can release a Switch 3 on the level of an XSS.

 

[UltraMagnus]

In 3-4 years they can release a Switch 3 on the level of an XSS.

Click to shrink...

So what is the point of having such a model in 2021 then? I just moves an actual Switch that could possibly handle more of those modern ports further down the pipeline and 1.5TFLOP is way overkill for just improving performance of current Switch tier games causing Nintendo to lower their profit margins (bad business) for no really great reason.

 

[Dekuman]

So what is the point of having such a model in 2021 then? I just moves an actual Switch that could possibly handle more of those modern ports further down the pipeline and 1.5TFLOP is way overkill for just improving performance of current Switch tier games causing Nintendo to lower their profit margins (bad business) for no really great reason.

Click to shrink...

The device being described is about what you can get for $300-400 in 2021. It's not that far off from the old speculation for a 2023 device that was around 2 to 2.5TF.

If next year's device is 1-1.5TF with DLSS it's basically the cutting edge device we all wanted.

I'm actually very surprised A78 cores are being thrown around as candidates for the 2021 device, as it was the standard case hope for the 2023 Switch 2 candidate in earlier discussions, though in those discussions the assumption was 8 cores.

We could still get 8 core A78s of course, and if what z0m3le is saying is true the A78C are for gaming and a cluster of those is 8 instead of 4, we could well get 8 core CPU that 2-3x more powerful that what's in a PS4 and places the Switch where it needs to be as a hybrid.

 

[Deleted member 37917]

In a prior podcast I've touched on the idea of a Lite connecting to a TV via a wireless docking solution similar to AirPlay.

Click to shrink...

I hadn't thought of it as a possibility. If this is the solution they choose, I believe that the flagship would also get it though.

If I had to guess Nintendo's strategy, I would guess that next year will be the flagship, and 2022 will be the upgraded lite model. I played charade tonight with friends and I won, so I'm kind of on a roll when it comes to guessing.

 

[UltraMagnus]

The device being described is about what you can get for $300-400 in 2021. It's not that far off from the old speculation for a 2023 device that was around 2 to 2.5TF.

If next year's device is 1-1.5TF with DLSS it's basically the cutting edge device we all wanted.

Click to shrink...

It needs to give Nintendo a fat profit margin on top of that, Nintendo is not here to be a charity to give people the top end mobile components for razor margins.

They probably enjoy a take of over $70+ profit on the base Switch right now and I don't think they're in any hurry to move away from that.

 

[Look over there]

So what is the point of having such a model in 2021 then? I just moves an actual Switch that could possibly handle more of those modern ports further down the pipeline and 1.5TFLOP is way overkill for just improving performance of current Switch tier games causing Nintendo to lower their profit margins (bad business) for no really great reason.

Click to shrink...

Do you think that the Switch as is can last until 2025?

 

[Dekuman]

It needs to give Nintendo a fat profit margin on top of that, Nintendo is not here to be a charity to give people the top end mobile components for razor margins.

They probably enjoy a take of over $70+ profit on the base Switch right now and I don't think they're in any hurry to move away from that.

Click to shrink...

Don't agree here, Switch did not give them a huge profit margin in 2017, i expect the 2021 will be the same. Breakeven/slight loss at launch then the margins ramp up.

 

[UltraMagnus]

Do you think that the Switch as is can last until 2025?

Click to shrink...

2025 as the primary driver, no but 2023, say? Sure.

We're already basically into 2021 and 2021 will have BOTW2, maybe Bayonetta 3, Monster Hunter Rise, and several other high profile games, so through 2021 they're in pretty good shape.

2022 they still have cards to play if they want like a price drop and a 2D Mario and things like that, but some decline in your 5th year on market is not something that is terrible. It's a normal part of the game hardware process.

 

[UltraMagnus]

Don't agree here, Switch did not give them a huge profit margin in 2017, i expect the 2021 will be the same. Breakeven/slight loss at launch then the margins ramp up.

Click to shrink...

Yeah but why would you be in a rush to go back to low profit margins? Like I'm sure someone on the board of directors when briefed on this would say "why exactly are we moving away from a $70-$80+ profit per unit that we can't even keep in stock to push some model that is sold near cost again?".

You better have a really, really good answer to that.

Enthusiasts are very centered on the concept of the game machine's performance, but if you're a business the *whole point* the whole dream of making a game console is to get to this point. The "harvest phase", where you are rewarded for the early years of work and now get to reap massive rewards and sky high profit margins.

For enthusiasts this part of the product cycle is probably more "boring" but for the company that's looking at it from a business POV, this is the part of the product cycle they want to get to from day 1.

 

[NineTailSage]

The 28nm, 16nm, etc. nodes of TSMC's competitors (mostly Samsung and GlobalFoundries) aren't compatible though. Switching from one manufacturer to another or even from one process to another usually requires a complete respin of the chip itself. Basically, once you've taped out a chip design with a particular manufacturer on a particular node of theirs, you're on that node and with that manufacturer unless you decide to "port" your design to another process and/or manufacturer. And this is not a simple process, but quite expensive and takes a least a few months. So TSMC shutting down what is still one of their biggest nodes (revenue-wise) would just leave a lot of their customers stranded (and pissed off) and it's not really good business sense for that reason alone. There are a lot of non-high-end chipsets being produced on those older nodes which work just fine and where there's absolutely no need to shrink them to a smaller node or anything.

Moreover, phasing out older nodes doesn't really mean more room for their 7nm and 5nm nodes. As far as I know, it's not really possible and/or feasible to upgrade older production lines to newer processes. To the best of my knowledge, these fabrication units have to be built pretty much from scratch every time and that's a multi-year endeavor anyway (just take a look at how Intel's manufacturing is pretty much fucked for at least a few years because their 10nm process didn't live up to their own expectations). So shutting down older processes would mean absolutely nothing wrt their capacity to deliver 7nm and 5nm chipsets in the short term. Yes, demand for various 7nm products far exceeds what TSMC can currently churn out, but there's really no short-term solution here. It really boils down to high-end semiconductor manufacturing being a seller's market at the moment and not a buyer's market.

Click to shrink...

Not saying that this will happen overnight but TSMC are definitely walking into the room and saying I'm charging this for these nodes and you will pay it or... 7nm, 5nm and 3nm TSMC will be able to charge quite a bit for awhile because there's just such a lack of competition.

 

[Deleted member 37917]

In 3-4 years they can release a Switch 3 on the level of an XSS.

Click to shrink...

2025 would be my guess i think a refresh every 4 years is a good balance. And it allows to keep a system alive for 5-6 years, which is just a bit less than a regular console generation.

 

[Dekuman]

Yeah but why would you be in a rush to go back to low profit margins? Like I'm sure someone on the board of directors when briefed on this would say "why exactly are we moving away from a $70-$80+ profit per unit that we can't even keep in stock to push some model that is sold near cost again?".

Click to shrink...

I think maintaning momentum is more important to Nintendo than squeezing profit. They tried that with the Wii and ended up waiting way too long with it.

I also agree with the earlier analysis that if Nintendo doesn't refresh next year, they risk turning the Switch into just a portable system that happens to have a TVout. It's appeal so far has been its ability to play current console and PC games. A more capable device next year will keep the flagship device at least in that space. Though I expect OG Switch will still be around for a while and maybe it'll morph into purely the 'lite' model completing its journey into what some people wanted out of Switch, a purely handheld device.

 

[UltraMagnus]

I think maintaning momentum is more important to Nintendo than squeezing profit. They tried that with the Wii and ended up waiting way too long with it.

I also agree with the earlier analysis that if Nintendo doesn't refresh next year, they risk turnin the Switch into just a portable system that happens to have a TVout. It's appeal so far has been its ability to play console and even some PC games. A more capable device next year will keep the flagship device at least in that space. Though I expect OG Switch will still be around for a while and maybe it'll morph into purely the 'lite' model completing its journey into what some people wanted out of Switch, a purely handheld device.

Click to shrink...

Again though the Switch is not really performing like the Wii at all. The Wii was already in notable decline at this point in its product cycle, the Switch is still accelerating sales wise.

The Wii's appeal was rooted in casual game fads and motion controller technology that was eventually copied by Sony/MS and then overshadowed by the rise of iPhone/Android gaming.

The Switch is not reliant on that at all. There is no real Wii Sports on the Switch, the Switch doesn't need it. Ring Fit is there, but the Switch was already a smash success way before that ever came out.

Sony/MS cannot copy the Switch so easily, they would have to invest fully into supporting an entirely new ecosystem and we saw Sony couldn't do that with the Vita, trying to offer a Switch tier competitor while also supporting the PS5 is a non-starter.

 

[Dekuman]

Again though the Switch is not really performing like the Wii at all. The Wii was already in notable decline at this point in its product cycle, the Switch is still accelerating sales wise.

The Wii's appeal was rooted in casual game fads and motion controller technology that was eventually copied by Sony/MS and then overshadowed by the rise of iPhone/Android gaming.

The Switch is not reliant on that at all. There is no real Wii Sports on the Switch, the Switch doesn't need it. Ring Fit is there, but the Switch was already a smash success way before that ever came out.

Sony/MS cannot copy the Switch so easily, they would have to invest fully into supporting an entirely new ecosystem and we saw Sony couldn't do that with the Vita, trying to offer a Switch tier competitor while also supporting the PS5 is a non-starter.

Click to shrink...

it's correct they appel to diffrent markets and audiences, but as i've already said part of Switch's appeal is its hybrid nature. People are still being blown away by the recent Crysis Remastered or Doom Eternal port on Switch. And Switch is at the table on service games like Warframe and Fortnite. (among others)

No refresh will mean those will stop happening and I have always questioned how long Epic and Digital Extremes and others can keep their service games on the Switch as those games evolve. A refresh makes sense financially from that perspective too, because I have no doubt games like Fortnite and Warframe bring lots of eshop purchases for skins and in-game currency where a cut goes to Nintendo, for literally doing nothing other than having those games on their platform. This is how Sony and MS is making loads of cash (in addition to their online subs). That's way more important than hardware margin these days.

 

[z0m3le]

1.5 TFLOPS portable with 50-68GB memory bandwidth that probably has to cost $350-$400 I think will just run into the same "problem" that it's supposed to be solving ... in 2 years (probably even less) many devs will say that's not good enough for modern 3rd party ports as they move into PS5/XSX specific development.

It also needs to run in the 11-12 watt range too, maybe 15 watts max like the current Switch does, things like 17-18 watts just for the GPU is way too high.

Click to shrink...

That 17-18 watt number isn't real. There is a RTX 3070 GPU that has 20TFLOPs and is only 220 watts. Even just slicing that GPU into a 10th is 2TFLOPs at 22 watts? That isn't how any of this works, and it is still showing that 1.5TFLOPs wouldn't be 3/4th the power consumption.

I'll make it clearer, the above includes GDDR6 memory, it's just someone making up a power number because it makes their argument sound better. Reality is that 1.5TFLOPs Ampere would consume no more than 10 watts. Heck there is a 1.1TFLOPs 14nm (Samsung) Pascal GPU called the GT 1030 which would out perform the 1.5TFLOPs Ampere GPU on the 8nm node.

The thread doesn't need bogus claims about power consumption sticking around for pages and being sited months from now. It's wrong, let it go.

 

[UltraMagnus]

it's correct they appel to diffrent markets and audiences, but as i've already said part of Switch's appeal is its hybrid nature. People are still being blown away by the recent Crysis Remastered or Doom Eternal port on Switch. And Switch is at the table on service games like Warframe and Fortnite. (among others)

No refresh will mean those will stop happening and I have always questioned how long Epic and Digital Extremes and others can keep their service games on the Switch as those games evolve. A refresh makes sense financially from that perspective too, because I have no doubt games like Fortnite and Warframe bring lots of eshop purchases for skins and in-game currency where a cut goes to Nintendo, for literally doing nothing other than having those games on their platform. This is how Sony and MS is making loads of cash (in addition to their online subs). That's way more important than hardware margin these days.

Click to shrink...

You can have a refresh without necessarily having to have a cutting edge new chip that lowers profit margins significantly.

To be honest if they just even let the existing Mariko run at full potential, that would give a lot of games a sizable boost.

 

[Brofield]

Is DLSS no longer on the menu for a Switch successor?

 

[ILikeFeet]

Is DLSS no longer on the menu for a Switch successor?

Click to shrink...

might still be on the menu

granted, we have not seen the menu yet

 

[Skittzo]

Is DLSS no longer on the menu for a Switch successor?

Click to shrink...

On the contrary it seems to be more likely based on Nate's and Thugstas' recent posts.

 

[Look over there]

2025 as the primary driver, no but 2023, say? Sure.

We're already basically into 2021 and 2021 will have BOTW2, maybe Bayonetta 3, Monster Hunter Rise, and several other high profile games, so through 2021 they're in pretty good shape.

2022 they still have cards to play if they want like a price drop and a 2D Mario and things like that, but some decline in your 5th year on market is not something that is terrible. It's a normal part of the game hardware process.

Click to shrink...

I think that 2023 won't allow for a Switch to match XSS. I agree with zakatana on 2025 to be the approximate time. My reasoning is that you need the generation after 3 nm to be up and running to soak up the big smartphone demand. Also, you'll need some time for 3 nm's price to depreciate. And 3 nm is the generation I have pegged to allow 8 ARM cores to comfortably run at low to mid 2 ghz within a few watts to match (or slightly beat even) PS5/XSS's CPU. Also, 4 years should give us another 2 generations of Nvidia GPU tech.

2023 gets you.... probably 5LPE as your best realistic option? Assuming we're sticking with Samsung here. And ARM projects their 2022 architecture, Matterhorn, to have about +30% performance over the A78. Ok, here's my rough math, which may or may not be terribly wrong. By all means, correct me where I'm off.
Starting with our power envelope allowing for a base figure of ~1.5 ghz for 8 A78 cores at 8LPP...
10LPE->8LPP should be 10LPE->10LPP (-15% power), then 10LPP->8LPP (-10%), thus 10LPE->8LPP ~= -23.5% power, or 0.765
10LPE->7LPP should be -50% power, or 0.5. Thus, the relation between the same work done on 8LPP and 7LPP is 0.765/0.5=1.53. Then to guestimate how much I can get out of investing the power saved back into clock rate, I take the sqrt of 1.53, which is about 1.23. Now multiply that by the original 1.5 ghz to get ~1.85 ghz. That is, I'm getting that 8 cores of (A78, or A78-like power efficiency) at ~1.85 ghz on 7LPP uses approximately the same power as 8 such cores at 1.5 ghz at 8LPP.

7LPP->5LPE should be -20% power, or 0.8. Flip it around to get 1.25, then take the sqrt, then multiply by ~1.85 to get 2.07.
Thus, my result is that 8 cores of (A78, or A78-like power efficiency) at 1.5 ghz on 8LPP, or 1.85 ghz on 7LPP, or 2.07 ghz on 5LPE all take roughly the same amount of power. Going with 5 LPE and assuming that Matterhorn offers +30% IPC over A78, you'd get the equivalent of about ~2.7 ghz A78. I don't think that's enough to match the PS5/XSS generation.

But again, my math could be all wrong here; my methodology comes from piecing together tidbits from here and there.

 

[UltraMagnus]

I think that 2023 won't allow for a Switch to match XSS. I agree with zakatana on 2025 to be the approximate time. My reasoning is that you need the generation after 3 nm to be up and running to soak up the big smartphone demand. Also, you'll need some time for 3 nm's price to depreciate. And 3 nm is the generation I have pegged to allow 8 ARM cores to comfortably run at low to mid 2 ghz within a few watts to match (or slightly beat even) PS5/XSS's CPU. Also, 4 years should give us another 2 generations of Nvidia GPU tech.

2023 gets you.... probably 5LPE as your best realistic option? Assuming we're sticking with Samsung here. And ARM projects their 2022 architecture, Matterhorn, to have about +30% performance over the A78. Ok, here's my rough math, which may or may not be terribly wrong. By all means, correct me where I'm off.
Starting with our power envelope allowing for a base figure of ~1.5 ghz for 8 A78 cores at 8LPP...
10LPE->8LPP should be 10LPE->10LPP (-15% power), then 10LPP->8LPP (-10%), thus 10LPE->8LPP ~= -23.5% power, or 0.765
10LPE->7LPP should be -50% power, or 0.5. Thus, the relation between the same work done on 8LPP and 7LPP is 0.765/0.5=1.53. Then to guestimate how much I can get out of investing the power saved back into clock rate, I take the sqrt of 1.53, which is about 1.23. Now multiply that by the original 1.5 ghz to get ~1.85 ghz. That is, I'm getting that 8 cores of (A78, or A78-like power efficiency) at ~1.85 ghz on 7LPP uses approximately the same power as 8 such cores at 1.5 ghz at 8LPP.

7LPP->5LPE should be -20% power, or 0.8. Flip it around to get 1.25, then take the sqrt, then multiply by ~1.85 to get 2.07.
Thus, my result is that 8 cores of (A78, or A78-like power efficiency) at 1.5 ghz on 8LPP, or 1.85 ghz on 7LPP, or 2.07 ghz on 5LPE all take roughly the same amount of power. Going with 5 LPE and assuming that Matterhorn offers +30% IPC over A78, you'd get the equivalent of about ~2.7 ghz A78. I don't think that's enough to match the PS5/XSS generation.

But again, my math could be all wrong here; my methodology comes from piecing together tidbits from here and there.

Click to shrink...

It also depends on what Nintendo wants and Nintendo is notoriously conservative. As much as some people may scoff at the notion, Nintendo may actually be fully OK with even just a moderate upgrade from what they have now all the way through 2025.

Don't assume ever that Nintendo is in the same head space as where hardcore gamers and technology enthusiasts are at.

When was the last time Nintendo released a system that had specs in line with what enthusiast gamers wanted? Having followed them for a long time, I think it was 27+ years ago when they revealed "Project: Reality" as their new console and the specs at that time knocked people's socks off (this was the N64).

Of course we didn't know then that they would also cripple the machine by not giving it a CD drive.

In Nintendo's case I think it's also better to look at hardware from the POV of "what are their hardware goals here" more than "what's the best possible technology they can get".

If the goal is to improve resolution/frame rate performance of the more challenging games on the system and even if the goal is to allow some PS4-tier games to be ported (alienating OG Switch owners be damned) to give the product line extra juice in the other half of the product cycle, you really don't need a huge upgrade here.

 

[ILikeFeet]

the very minimum of NIntendo's goals probably lie in removing the bottlenecks, so improving the cpu and memory bandwidth. doing just that, I'd say the worst they could do is 4 A76s @ 1GHz, 4GB LPDDR4X for 50GB/s, and 4 Ampere SM for 256 fp32 cores (and 256 fp32/int32 cores) at the same clocks as the OG. for DLSS, I'd question if that's enough to apply DLSS to all possible titles, including the 60fps ones.

 

[z0m3le]

the very minimum of NIntendo's goals probably lie in removing the bottlenecks, so improving the cpu and memory bandwidth. doing just that, I'd say the worst they could do is 4 A76s @ 1GHz, 4GB LPDDR4X for 50GB/s, and 4 Ampere SM for 256 fp32 cores (and 256 fp32/int32 cores) at the same clocks as the OG. for DLSS, I'd question if that's enough to apply DLSS to all possible titles, including the 60fps ones.

Click to shrink...

Wouldn't be able to run Switch games as well as the Switch. That GPU is actually less powerful.

 

[ILikeFeet]

Wouldn't be able to run Switch games as well as the Switch. That GPU is actually less powerful.

Click to shrink...

I figured having the same amount of pure fp32 cores would be suffice. unless Ampere changed up how they worked considerably than I realized