StarTopic Future Nintendo Hardware & Technology Speculation & Discussion |ST| (New Staff Post, Please read)

[Νεφέλωμα]

What if each "Joycon 2" had two grips to conserve horizontal symmetry for standalone play? Then in handheld mode you would only use the grip at the bottom. Thoughts?

 

[D36]

What happened while I was asleep???? This thread was shmovin

 

[eye0h]

About

We are The Nintendo Pipeline! We are a group of passionate Nintendo fans who are always on top of the latest news and games. This community blog is where you’ll see original content created a…

nintendopipeline.wordpress.com

damn, discord link is dead :/

 

[Ghostsonplanets]

and I have done some back of the envelope calculations, and I don’t think that DLSS 3 would be viable regardless of the architecture. Just not fast enough on a little machine

Could you provide the link for it Oldpuck? I must have missed it. But it's not surprising. The GA10F OFA is also the same as Ampere, right?

 

[Catsylvania]

You mean.....sort of stretched...... like......
butter scraped over too much bread?

 

[Mr Swine]

Has anyone done any calculations on how much bandwidth 1MB of L2 does (2MB, 4MB, 8MB and so on) on Ampere/Lovelace?

Looking at the 4060 Ti on a 128bit bus with 288GB/s bandwidth and 32MB L2 cache trades blows with the 3070 that has a 256bit bus with 448GB/s bandwidth.

Would be fun if Switch 2 could have a large L2 cache to mitigate the low bandwidth but that won’t happen

 

[Stefan51278]

I have them both, neither give me Nintendo games.

I'm not suggesting they "go back to home console" and reverse the hybrid success but I'd love them to have a "pro" box as a niche market in addition to Switch 2 hybrid.
It won't happen but I'd love it.

You should really go for the emulator-route if you want the beefier Switch Experience. I played Tears of the Kingdom in 4K with an emulator for some time and it works. The only catch is the beefier gaming-pc you need for this.

 

[mjayer]

NX2 isn't the code name for the next device, but NX is the name of the platform which includes the Switch successor, we know this because of the Nvidia leak, so NX2 is a good quick hand for it, though you might be explaining that it's just a community nickname for the Switch successor until we have a codename or official name. We also have plenty of other nicknames, such as the Succ, Switch 2, REDACTED, and NG Switch.

Saw some power profile thing being passed around for the "devkit" all rumor, but 11w portable and 28w+ when docked... It's important to remember that devkits will have a higher CPU clock to run the debug software and just for general stability, Switch devkits for instance run at 1.2GHz. This could count for 1 or 2 watts, so retail units would draw 9 to 10 watts when portable if this rumor is correct.

The docked number is also easily explained, to me this is a case of a 15w draw for charging the battery, thus the power draw for the system is 13w to 14w, again with 1 to 2 watts going towards that higher CPU clock, basically these numbers more or less match launch Switch units, another thing to consider is that 5000mah batteries are very common in both high end and budget phones, this is a reasonable upgrade to expect for Switch's successor, as the Switch used a 4315mah battery.


Drake/T239 uses GA10F (Geforce/Graphics Ampere) it is custom and seems to have some upgraded components, which includes the optical engine that Orin uses, that is vital to DLSS 3.0's frame generation, however it's separate from Ada Lovelace, and we don't know which is superior or if Orin/Drake's Optical engine is fast enough to support frame generation. We will learn if it is at a later time.

Entertaining the slight possibility this leak is real, for the fun of it...

With this power draw what are the power levels, can you do your amazing analysis for us? Would we be looking at a dreamy 3.X Teraflop Switch 2? Also portable TF levels? Was he right in saying (with those numbers) it is comparable to a Series S, ish?

 

[Deleted member 887]

Could you provide the link for it Oldpuck? I must have missed it. But it's not surprising. The GA10F OFA is also the same as Ampere, right?

The OFAs appear the same, yeah.

Thraktor starts the discussion here, and we go back and forth in a few replies. But I can summarize.

DLSS 3 frame gen has two phases, to hugely over-simplify. The first is to analyze the two frames it has buffered. The second is to generate a new frame between them. The OFA is critical for that first phase, but the second phase happens on some combination of tensor/shader cores.

It's not clear how much of the process is one phase or the other, but now that there are some 4060 benchmarks out there, we can start to take a guess. And that guess says that Frame Gen could take as long as 40ms on on a 3 TFLOPS machine. At that point it's taking longer to make your "fake" frame than it would to just generate a new one natively.

At that point, the OFA from Ada is actually a liability. It's like 3x larger, physically, so that's a decent chunk of wasted space on the chip which could be spent elsewhere. Not huge, really, because the OFA is probably pretty small, but for a mobile chip, every bit counts.

 

[Foltzie]

Has anyone done any calculations on how much bandwidth 1MB of L2 does (2MB, 4MB, 8MB and so on) on Ampere/Lovelace?

Looking at the 4060 Ti on a 128bit bus with 288GB/s bandwidth and 32MB L2 cache trades blows with the 3070 that has a 256bit bus with 448GB/s bandwidth.

Would be fun if Switch 2 could have a large L2 cache to mitigate the low bandwidth but that won’t happen

I would have to defer to @oldpuck and @Thraktor for any simulations, but since T239 has been developed for Nintendo specifically whatever L2 amount they have chosen should be more than enough to make sure Redacted isn't memory bandwidth limited.

Older posts in this thread indicated L2 is either 1MB or 4MB in the NVN2 leak.

 

[Hermii]

I would have to defer to @oldpuck and @Thraktor for any simulations, but since T239 has been developed for Nintendo specifically whatever L2 amount they have chosen should be more than enough to make sure Redacted isn't memory bandwidth limited.

Older posts in this thread indicated L2 is either 1MB or 4MB in the NVN2 leak.

Cost is also a huge issue for Nintendo though. They're not trying to reduce bottlenecks at all costs, they're trying to be the best they can be within budget.

 

[Armoured_Bear]

You should really go for the emulator-route if you want the beefier Switch Experience. I played Tears of the Kingdom in 4K with an emulator for some time and it works. The only catch is the beefier gaming-pc you need for this.

Thanks a lot but I'm finished with Gaming PCs.
I built my own in a HTPC case for the living room and used it as a 3rd party machine for a good few years.
I've completely given up on it (it's now old anyway), I spend all day working in Software development and have little interest nowadays in endless tinkering and fixing stuff when gaming.
No matter what people say, PCs with Steam etc. are not plug and play, every so often there are issues, so many shitty ports, tinkering with all settings to optimise FPS etc. If people enjoy that then fair play to them but as a middle-aged guy, I want to switch my console on and play.

I did mess around with a Wii or GC emulator some years ago but it's really not for me, as enticing as playing ToTK in 4K is, I hope to get it with Switch 2.

 

[Zemya]

How much usually does it take from devkits to console reveal?

 

[Skittzo]

How much usually does it take from devkits to console reveal?

Devkits go out in a number of phases. We don't know what phase these supposed devkits are. If they're initial devkits we could still be looking at a year+, if they're final devkits then we could be looking at weeks before reveal.

 

[ILikeFeet]

How much usually does it take from devkits to console reveal?

there's no "usually". it's whenever the timing permits

 

[Concernt]

All respect to Rich, but he hasn't studied the Nvidia leak.

It has backported a couple of power saving features, and may be on 5/4nm but other that it's ampere.

Better node, better OFA and power saving features, that's the majority of Lovelace's advantages over Ampere.

 

[Hocot8]

Would someone working on Fortnite know Nintendos first party plans? Would they be dumb enough to say things only someone working on Fortnite would know? Would the BC stuff even be relevant to their work?

Imo it's 100% fake and not even worth entertaining.

I'm slightly confused is this the same fortnite rumor that came out several months ago?

I really think the main new thing of the Switch 2 is gonna be new controllers. "Joycons 2" maybe. I would love the next detachable controller to just be a pro controller split in half, like some fan renders show. Also, I think the console itself is gonna have a more ergonomic design, instead of being just a tablet.

I think a more rounded approach is in the cards. The joycons, and I think most agree, can be a little bit bigger.

 

[Hermii]

Better node, better OFA and power saving features, that's the majority of Lovelace's advantages over Ampere.

Except it has the same OFA as desktop ampere.

 

[Ghostsonplanets]

The OFAs appear the same, yeah.

Thraktor starts the discussion here, and we go back and forth in a few replies. But I can summarize.

DLSS 3 frame gen has two phases, to hugely over-simplify. The first is to analyze the two frames it has buffered. The second is to generate a new frame between them. The OFA is critical for that first phase, but the second phase happens on some combination of tensor/shader cores.

It's not clear how much of the process is one phase or the other, but now that there are some 4060 benchmarks out there, we can start to take a guess. And that guess says that Frame Gen could take as long as 40ms on on a 3 TFLOPS machine. At that point it's taking longer to make your "fake" frame than it would to just generate a new one natively.

At that point, the OFA from Ada is actually a liability. It's like 3x larger, physically, so that's a decent chunk of wasted space on the chip which could be spent elsewhere. Not huge, really, because the OFA is probably pretty small, but for a mobile chip, every bit counts.

Thank you!

Better node, better OFA and power saving features, that's the majority of Lovelace's advantages over Ampere.

Also better media engines.

 

[Foltzie]

Cost is also a huge issue for Nintendo though. They're not trying to reduce bottlenecks at all costs, they're trying to be the best they can be within budget.

Sure, but other than Switch itself, Nintendo's consoles have rarely been limited on RAM or memory bandwidth. Whatever Redacted has should be sufficient to keep the CPU and GPU feed.

1MB of L2 seems low to me, but w/o more context (like a comparison of L2 levels to RAM and memory bandwidth configurations on a modern SOC) I can only guess.

 

[D36]

About

We are The Nintendo Pipeline! We are a group of passionate Nintendo fans who are always on top of the latest news and games. This community blog is where you’ll see original content created a…

nintendopipeline.wordpress.com

The Discord link isn't working for me.

 

[Cybergatuno]

Do we still think Nintendo must rush out a successor?

 

[Skittzo]

Do we still think Nintendo must rush out a successor?

I think, and have thunk for years, that current sales trajectory has absolutely no bearing on the development and release of a successor.

 

[Hermii]

Sure, but other than Switch itself, Nintendo's consoles have rarely been limited on RAM or memory bandwidth. Whatever Redacted has should be sufficient to keep the CPU and GPU feed.

1MB of L2 seems low to me, but w/o more context (like a comparison of L2 levels to RAM and memory bandwidth configurations on a modern SOC) I can only guess.

I can think of a console that's limited by memory bandwidth, it's called the Nintendo Switch.

Memory bandwidth is an issue for portable tech in general though, due to the high power consumption of fast memory.

 

[Concernt]

Except it has the same OFA as desktop ampere.

We don't exactly know that. What's much more likely is that it inherits the OFA from the existing ORIN devices (which don't have exactly the "same OFA as desktop ampere").

 

[SpringwoodSlasher]

New intel from Nikki

Same Horizon OS as Switch 1 for the NG


Confirms how the Switch 2 chip performances will be reduced compared to the retail Nvidia Orin chip

So I’m clearly out of the loop these past few months.

I’m assuming general speculation has been leading towards a Switch 2 rather than a new idea/gimmick for a console. Would I be correct in assuming that?

 

[Hermii]

We don't exactly know that. What's much more likely is that it inherits the OFA from the existing ORIN devices (which don't have exactly the "same OFA as desktop ampere").

Ilikefeet and Oldpuck just contradicted you.

same OFA as desktop

The OFAs appear the same, yeah.

Thraktor starts the discussion here, and we go back and forth in a few replies. But I can summarize.

DLSS 3 frame gen has two phases, to hugely over-simplify. The first is to analyze the two frames it has buffered. The second is to generate a new frame between them. The OFA is critical for that first phase, but the second phase happens on some combination of tensor/shader cores.

It's not clear how much of the process is one phase or the other, but now that there are some 4060 benchmarks out there, we can start to take a guess. And that guess says that Frame Gen could take as long as 40ms on on a 3 TFLOPS machine. At that point it's taking longer to make your "fake" frame than it would to just generate a new one natively.

At that point, the OFA from Ada is actually a liability. It's like 3x larger, physically, so that's a decent chunk of wasted space on the chip which could be spent elsewhere. Not huge, really, because the OFA is probably pretty small, but for a mobile chip, every bit counts.

 

[Concernt]

Ilikefeet and Oldpuck just contradicted you.

Note the word "exactly".

 

[KeijiKG]

Is this thread even about hardware anymore? Is this the new Paper Mario thread?

No.

 

[Deleted member 887]

Has anyone done any calculations on how much bandwidth 1MB of L2 does (2MB, 4MB, 8MB and so on) on Ampere/Lovelace?

Looking at the 4060 Ti on a 128bit bus with 288GB/s bandwidth and 32MB L2 cache trades blows with the 3070 that has a 256bit bus with 448GB/s bandwidth.

Would be fun if Switch 2 could have a large L2 cache to mitigate the low bandwidth but that won’t happen

The short version is, "I would expect 1MB of L2 to perform roughly like other RTX 30 cards."

The longer version is, you can't really simulate the effect without knowing the cache hit rate, but we can compare to other GPUs in the same architecture. To keep it simple, I'm going with full GPU dies, not all the binned variants.

I'm assuming the smaller 1MB L2 cache here, and a lot of this analysis is me just summarizing @Look over there's very smart work.

GPU	Bandwidth/TFLOP	Cache/Bandwidth
GA102	25 GB/s/TFLOP	6KB
GA103	22.4 GB/s/TFLOP	9KB
GA104	28.95/GB/s/TFLOP	7KB
GA106	30GB/TFLOP	8KB
GA107	24GB/TFLOP	9KB
T239 (3TFLOPS, docked)	34GB/TFLOP	10KB
T239 (1.5TFLOPS portable)	44GB/TFLOP	15KB

You can see that T239 has more bandwidth than any other card in the RTX 30 range. Only the weirdly large GA106 comes close. Now, ideally, a console needs more bandwidth than a standard GPU, but the bandwidth is shared with the CPU, but this looks pretty good.

And you can see that the cache is also slightly on the high end, even with just a piddly 1MB. Considering how big caches are, a 1MB design (instead of the 4MB we've seen elsewhere) makes a ton of sense.

 

[Hermii]

Note the word "exactly".

It's a pointless discussion anyway. Even if it was Lovelace OFA frame generation would be useless on Drakes power envelope.

 

[Skittzo]

So I’m clearly out of the loop these past few months.

I’m assuming general speculation has been leading towards a Switch 2 rather than a new idea/gimmick for a console. Would I be correct in assuming that?

Nobody knows really. It's safe to assume that it'll be similar in form factor to the Switch based on the chip they're using and the idea that they want to carry over Nintendo accounts/supposedly the OS seems to suggest an iteration of the Switch concept rather than something brand new.

But we really don't know for sure at this stage, it could very easily be similar to the Switch but also include something radically new and different. All we really "know" is the SoC.

 

[Yavga]

I'm still waiting for the Nvidia Shield Tablet 2

Nubia is making gaming tablets, technically the same thing if what you need is a gaming orientated Android tablet. They also have a 3d tablet or something.

The price however is not the same...

 

[Concernt]

It's a pointless discussion anyway. Even if it was Lovelace OFA frame generation would be useless on Drakes power envelope.

Oh, absolutely, yes. At least we can agree there. I'm definitely not getting on DLSS 3.X, at least not the current form of it.

I might speculate at the possibility of non-DLSS frame generation, if enough headroom can be found. I'm fairly sure T239 wouldn't cross the line as regards making FSR3 work acceptably, though I could be wrong. Still, technically, interpolation and interlacing is a trick available to developers who have a game working fine EXCEPT for framerate.

(An example on Switch 1 would be, Mario Odyssey and Bowser's Fury both produce the same amount of pixels per second in handheld mode barring lag, but Odyssey uses interlacing to achieve a 720i60FPS presentation, while Bowser's Fury uses 720p30.

Personally I prefer a cleaned up interlaced image to a lower framerate progressive image, but that isn't always an option.)

 

[Link_enfant]

Do we still think Nintendo must rush out a successor?

Please tell me this is being ironic. Reading anything about Nintendo "rushing out a successor" about 6.5 years after the Switch's release can't be serious, and it tends to hugely trigger me lol

 

[Mirev]

Do we still think Nintendo must rush out a successor?

The demand is high, it’s time to strike

 

[Gotdatmoneyy]

Do we still think Nintendo must rush out a successor?

Well it isn't coming out this year and it isn't coming out in Q1 of next year so at that point, a fall 2024 release isn't really rushing. 7-7.5 years on the market is about the standard PS/Xbox generation and way longer than the normal Nintendo one.

Also, even if Switch is up in Japan for the quarter YoY (I actually expect it'll be up everywhere YoY for the Q) it still is going to decline from the previous fiscal year when all is said and done. Holiday 2024 is the period and 1 or 2 months of exceptional performance are not going to stave off the incoming decline.

Don't let good results confuse reality.

 

[Giancarlo]

I'm slightly confused is this the same fortnite rumor that came out several months ago?


I think a more rounded approach is in the cards. The joycons, and I think most agree, can be a little bit bigger.

if Nintendo has decided to keep the Joy-Con or something similar for the Switch sucessor, i hope they do something to solve the dreadful Joy-Con drift, a lot of my friends suffered trough this problem

 

[Discostew]

Got in late for the HOS discussion, but I don't think we're going to see them use the exact same HOS used by Switch. More like HOS 2.0, where the baseline for it involves its security. Even though Switch got hacked early on, it was because of the convenience of the entry point being well documented and public from the situation with the Shield TV. Hackers had said that if it wasn't for that point of entry, they would still be looking for another opening, perhaps even now, because they felt that HOS was that secure.

The way HOS works with memory is it partitions it into 4 sections - Application, Applet, System and System Unsafe. The latter two are, imo, the basic portions of the OS, which together allocate roughly 310MB. Application is for the games, taking ~3.2GB. The remaining 467MB is for Applet, which is everything else, like the Home Menu, Album, eShop, Settings, even the All Software list etc, and as you move into each part of the system, the prior stuff is emptied out and newer stuff is loaded in. Areas like the eShop and Album are loading in data as you scroll through them, but aren't retained because there's only a limited amount of memory to work with, so when you scroll back and forth, it's continually reloading. For the eShop, the data isn't located locally, so it has to continually download. While folks suggest that the reason Nintendo didn't include a fully loaded internet browser was to avoid a hacking situation, it could also be because of the limited amount of RAM that would have been available to it. Can't really compare it to the Wii U, because as time goes on, there's expectations from things like that, like using the latest software, and the available RAM likely couldn't cut it.

For the eShop, the CPU is used heavily, using the 3 game cores when a game isn't loaded, but is limited to the OS core when a game is loaded (the latter of these two scenarios brings the eShop to a crawl). Back to the point of the eShop redownloading data as you scroll, this is likely why the eShop uses a good amount of CPU. I think it's having to process this data, which may include decompression. Let's say it is having to decompress data for the eShop. With Switch 2, if they redesign the eShop, they could possibly make use of the FDE so the CPU isn't being pushed, which could make browsing the eShop much faster (besides having the stronger CPU).

What I'm getting at ultimately, is let's not think of Switch 2 using Switch's HOS "as-is", but as a baseline for the important things while also being expandable in content, features, and functionality.

 

[Giancarlo]

Please tell me this is being ironic. Reading anything about Nintendo "rushing out a successor" about 6.5 years after the Switch's release can't be serious, and it tends to hugely trigger me lol

7/8 years for the past two console generation has been the typical lifecycle of a console, if Nintendo Switch sucessor launch in holiday 2024 as i believe Nintendo will do with the console, that would be longest console generation for Nintendo, Switch would be 7 years old and 8 in the market, sufficient time for a sucessor to launch, here a intereting comparison of previous Nintendo console games on it seventh year on the market compared to Switch seventh year

How Does Switch's Year 7 Compare To Previous Nintendo Consoles?

The magnificent seven?

www.nintendolife.com

 

[chocolate_supra]

Do we still think Nintendo must rush out a successor?

It's been covered a lot in this thread by people much smarter about this stuff than I am, so I'm just regurgitating what I've read here over the years:

Hardware development takes so much time, so much scheduling, so much coordination between multiple vendors, foundries, etc, that they cannot and would not plan their release around how well sales are going. They cannot pivot on a dime like that to hold off on production or sit on the upcoming system if current sales are high, then suddenly release it now once sales start plateauing. Whenever the succ releases, that will have been the plan for a long time with very little wiggle room to change it up.

The strength of sales of either consoles or games should not be a factor when trying to determine hardware release schedules because there's no way Nintendo could predict such things years ahead of time, which is when they'd be laying out schedules and reserving foundry space.

 

[MP!]

Got in late for the HOS discussion, but I don't think we're going to see them use the exact same HOS used by Switch. More like HOS 2.0, where the baseline for it involves its security. Even though Switch got hacked early on, it was because of the convenience of the entry point being well documented and public from the situation with the Shield TV. Hackers had said that if it wasn't for that point of entry, they would still be looking for another opening, perhaps even now, because they felt that HOS was that secure.

The way HOS works with memory is it partitions it into 4 sections - Application, Applet, System and System Unsafe. The latter two are, imo, the basic portions of the OS, which together allocate roughly 310MB. Application is for the games, taking ~3.2GB. The remaining 467MB is for Applet, which is everything else, like the Home Menu, Album, eShop, Settings, even the All Software list etc, and as you move into each part of the system, the prior stuff is emptied out and newer stuff is loaded in. Areas like the eShop and Album are loading in data as you scroll through them, but aren't retained because there's only a limited amount of memory to work with, so when you scroll back and forth, it's continually reloading. For the eShop, the data isn't located locally, so it has to continually download. While folks suggest that the reason Nintendo didn't include a fully loaded internet browser was to avoid a hacking situation, it could also be because of the limited amount of RAM that would have been available to it. Can't really compare it to the Wii U, because as time goes on, there's expectations from things like that, like using the latest software, and the available RAM likely couldn't cut it.

For the eShop, the CPU is used heavily, using the 3 game cores when a game isn't loaded, but is limited to the OS core when a game is loaded (the latter of these two scenarios brings the eShop to a crawl). Back to the point of the eShop redownloading data as you scroll, this is likely why the eShop uses a good amount of CPU. I think it's having to process this data, which may include decompression. Let's say it is having to decompress data for the eShop. With Switch 2, if they redesign the eShop, they could possibly make use of the FDE so the CPU isn't being pushed, which could make browsing the eShop much faster (besides having the stronger CPU).

What I'm getting at ultimately, is let's not think of Switch 2 using Switch's HOS "as-is", but as a baseline for the important things while also being expandable in content, features, and functionality.

Interesting post
I could tell that’s exactly what was happening in eShop but I wonder why there’s not a cache option to cache all the icons locally (internal storage space limits?)

 

[Maxrunner]

i thought the ts239 wouldnt support dlss3.

 

[D36]

You do realise that not a single person in this thread has suggested even for a second that Nintendo will abandon the hybrid model, don't you?

I don't think that's what they were implying.

 

[Mr Swine]

The short version is, "I would expect 1MB of L2 to perform roughly like other RTX 30 cards."

The longer version is, you can't really simulate the effect without knowing the cache hit rate, but we can compare to other GPUs in the same architecture. To keep it simple, I'm going with full GPU dies, not all the binned variants.

I'm assuming the smaller 1MB L2 cache here, and a lot of this analysis is me just summarizing @Look over there's very smart work.

GPU	Bandwidth/TFLOP	Cache/Bandwidth
GA102	25 GB/s/TFLOP	6KB
GA103	22.4 GB/s/TFLOP	9KB
GA104	28.95/GB/s/TFLOP	7KB
GA106	30GB/TFLOP	8KB
GA107	24GB/TFLOP	9KB
T239 (3TFLOPS, docked)	34GB/TFLOP	10KB
T239 (1.5TFLOPS portable)	44GB/TFLOP	15KB

You can see that T239 has more bandwidth than any other card in the RTX 30 range. Only the weirdly large GA106 comes close. Now, ideally, a console needs more bandwidth than a standard GPU, but the bandwidth is shared with the CPU, but this looks pretty good.

And you can see that the cache is also slightly on the high end, even with just a piddly 1MB. Considering how big caches are, a 1MB design (instead of the 4MB we've seen elsewhere) makes a ton of sense.

Thanks for taking your time, sounds good that it has higher bandwidth than the other Ampere cards. The CPU should take a chunk of that, in guessing it will be similar to what’s reported on the PS4 CPU taking a chunk out of the GDDR5 bandwidth?

 

[Deleted member 887]

So, sharing some back of the envelope math from yesterday that came up in a discussion with some DF folk. Back of the envelope is literal, actually, as I was doing a large computer migration at the time, and was scribbling on the back of a documentation packet to keep myself occupied.

GPU	Size	CUDA Cores	Transistors	Cost
GA102 (RTX 3090+)	628mm^2	10752	28,300	$1499
AD102 (RTX 4090+)	609mm^2	18432	76,300	$1599

What the heckfire does this mean, OldPuck?

It's about nodes again, sorry. The question is, what node will Nintendo use, and I think this thread has generally trended towards believing in 5nm as the only viable option for how big the damn GPU is. But I wanted to come at it from the other direction, which is cost. We know 5nm is expensive, but what does that actually mean in practice?

Looking at the cost numbers here, we can see that the RTX 3090 and 4090 launched at similar price points, but the 4090 was much larger in terms of transistors and CUDA cores, while being physically smaller. That's the trick with a node shrink. The wafers are more expensive, but you can fit more transistors into every square millimeter. You can spend that increase making chips more powerful at the same cost (thats how you get gen-on-gen performance upgrades) or the same level of power, but cheaper, and more energy efficient (that's how the mobile industry has come so far).

So the question is, what does 5nm cost per chip, assuming the chip stays the same size? This numbers roughly suggest a 2.7x density improvement. Not all arches shrink the same, and CUDA cores are slightly less than what you'd expect, but also RTX 4090 spends some of it's die space on making the OFA bigger. So let's go with the 2.7 number as a start - if 5nm is 2.7x as expensive per wafer as 8nm, then 5nm is break even in cost with 8nm*

How much do wafers cost? Well, there isn't official data, but I've seen some numbers float around. I was initially going to use those numbers, but it turns out they come from a paper written by some engineers at Georgetown, and the original paper is more useful than the numbers extracted from it.



As you can see, the prices are pretty consistent! The point the paper is trying to make is that the flattening of costs for production is being countered by reductions in power cost. For us that means that improvements in battery life. I've highlighted a few nodes, but let's pick out the 16/12 node.

In 2019, Nintendo launched the Lite and the revised Switch, this moved the node over to 16nm, which was 4 years old at the time. You'll see that come next year, the 5nm node will be just as old as 16nm was then, and started off more expensive for chips of the same transistor count. Adjusted for inflation, it's even more dramatic. That suggests that, in terms of cost, 5nm is in line with what Nintendo is willing to pay for better battery life, and that it's possible to offer a profitable consumer product at similar prices to 8nm.

I don't know what node Nintendo will choose, but I am increasingly convinced that cost isn't an issue for 5nm

*Assuming yields are the same

 

[Heron]

Regarding a Switch 2 home console which could be like a "Pro" model to the standard edition, there are two elements to consider.

One is whether it would be technically feasible for both hardware and software development. I'll leave the hardware part to others, but given that many 3rd parties are already using middleware like unreal or unity to create games for 6+ specs already (between PS4, PS4 Pro, PS5, XB1, XB1X, XSS, XSX, Switch Handheld, Switch Docked, PC etc) I can't imagine that's a problem. These middleware engines are specifically designed to make this process easy and studios are used to this workflow. Not a big consideration imo.

But the second and more important point is whether it makes business sense, which is where I think there are clear differences between the PS4 Pro model and what an NX2 home console would be. Those mid-gen refreshes ended up making up about, what, an estimated 20% of sales of those consoles from the point they were released (if someone has current figures on this let me know). Does that recoup the R&D costs from developing the consoles? I don't have the internal figures to be able to say.

But what I can say is that development of the consoles was almost certainly motivated by competition. As 4k TVs began to proliferate, Xbox had a chance to steal some of Sony's marketshare by having the only 4k box. Sony was at risk of losing image-quality conscious consumers to both Xbox and potentially to home-gaming PCs. So once one of the two companies began to walk down that path, the other had to follow to protect their user-base. From that perspective, and considering all the software sales that Sony probably made on PS4 Pro, I'd wager the R&D was worth it to them. There is probably a lot of corporate espionage going on between both companies right now to figure out whether each is making a Pro version of the current gen - because if so, the other will have to follow, but if not, it's better for them not to bother. (I note that Xbox in the FTC filing appears to believe that Sony IS making a Pro, which surprises me as there's no obvious reason like 4k support to make one.)

But does any of that apply to Nintendo? Do they have that kind of competition? No. Steam Deck, the ROG Ally etc... these are not actually feasible competition. Because the unique selling point of Nintendo hardware is Nintendo games, not graphically intensive 3rd party games. So sure, a Nintendo home console NX2 running a better chip at higher clock speeds (or what have you) might sell to 20% of their base... but they aren't at any risk of losing those consumers if they don't make that product. So the R&D cost of making it is a pure loss.

And would they even sell to 20% of their base? Whereas upgrading to a PS4 Pro from PS4 required no sacrifice and brought only benefit to the person willing to shell out the money, to buy a home console instead of a Switch requires giving up the handheld ability. Given the stats we've seen from Nintendo on how often people use handheld mode at least some of the time, I imagine that would bring the appeal down to less than that 20% of people who previously bought Pro consoles from the other manufacturers. Many of those purchasers would be buying in ADDITION to a regular hybrid - which means no extra software sales (in contrast to the Switch Lite, which was designed to grow the userbase in family households and thus increase software sales).

I am actually one of those people who want this. I would give up the handheld ability for even a slightly stronger box. But Nintendo knows they've got my sales either way, because I can't play their games anywhere else. So they'll never cater to me in this way. I've made peace with that - but I will continue to whine from time to time!

One last point - when the Switch was revealed I remember saying to other Nintendo fans online among all the celebrating "You realize that this is Nintendo abandoning home consoles for good, right? Shouldn't we at least have a retrospective or something, send them off properly?" And was promptly told that I was an idiot, that Switch meant nothing of the sort, third pillars etc. Now, 6 years later it's very obvious that they're never going back, yet we never really had that moment where we said goodbye to Nintendo's home consoles as a distinct thing. I guess there's always a chance they return one day, but...

 

[eye0h]

Do we still think Nintendo must rush out a successor?

Japan are just buying them for sport at this point jeez

 

[ReddDreadtheLead]

I'm a little confused tbh, is the T239 fully Ampere with some lovelace features or is it some sort of bogged down lovelace *SoC? Does this completely dismiss DLSS 3 for REDACTED? 'pologies if I missed a post explaining it reading through the trainwreck ,:v
(also my first famiboards post hi o/)

It’s Ampere with some aspects of Lovelace, but for all intents and purposes for gaming it is Ampere. Plus, at the size and the constraints that device would operate in, DLSS3 probably isn’t a good idea.

"I don't think anyone has dug into how cut down"?

What the hell has this thread been doing all this time?

And unless she knows something we don't, it's far less cut down than people would guess.

Idk, it seems pretty cut down.

Has reduced SM/CUDA Core count, has reduced CPU cores, has significantly reduced L2 cache, has PVA and DLA removed (Orin has 2), has reduced memory interface (128b vs 256b), has reduced tensor core performance and reduced FP16 performance for the Ampere CUDA cores, doing 1:1 instead of 1:2 performance. Does 48Warps instead of 64Warps, etc.

That’s already really cut down, T239 is focused for gaming of course, but it’s already really cut down in many area compared to ORIN.

During the FTC trial, there were claims that half the PS5 owners also owned a Switch.

*Almost Half of the PS5 owners in the United States.

I’d be happy just with more theme colours than just black or white

I’d like an actual black for the OLED at least, this is just Grey .

All respect to Rich, but he hasn't studied the Nvidia leak.

It has backported a couple of power saving features, and may be on 5/4nm but other that it's ampere.

Ada is already largely Ampere by 92% of what it is.

Do we know Orin or Drake has a different OFA from ampere?

same OFA as desktop

The OFA in ORIN is different from the desktop, and T239 has the same one as ORIN. Due to ORIN’s usecase (automotive, object detection) it’s critical that it has a very performant and adequate OFA especially for this bolded reason. Desktop Ampere is only able to deliver for video playback, but ORIN’s OFA is suitable for real-time object detection.

That said, I wouldn’t expect DLSS3 as the envelope this device runs at is seemingly way too small and way too low, but maybe if the hardware being a set console and not PC it could be different? It’s hard to say as DLSS FG is still an area that needs a lot of study.

Has anyone done any calculations on how much bandwidth 1MB of L2 does (2MB, 4MB, 8MB and so on) on Ampere/Lovelace?

We can’t figure that out unless we know how much latency the L2 has. Or the clock frequency as it’s tied to the GPU.


But I’m advocating for an SLC, as it helps reduce power consumption but also acts to reduce power consumption. Plus, all ARM SOCs since 2017 have an SLC which gives them coherency.

Including NVidia’s ones, though the way they did it with Xavier had weird quirks to it and they ended up abandoning that for the one ARM provides with the ORIN SOCs having the 4MB of System Cache. In any case for Xavier, that had the L3 take place as the LLC/SLC, although buggy.

i thought the ts239 wouldnt support dlss3.

It won’t.

Thanks for taking your time, sounds good that it has higher bandwidth than the other Ampere cards. The CPU should take a chunk of that, in guessing it will be similar to what’s reported on the PS4 CPU taking a chunk out of the GDDR5 bandwidth?

Shouldn’t be more than 25GB/s. Probably about 20GB/s. It’s up to how they choose to develop a game though.

I don't know what node Nintendo will choose, but I am increasingly convinced that cost isn't an issue for 5nm

*Assuming yields are the same

At that size, yields also aren’t much of an issue. . Even if Nintendo goes sparser than 125MTr/mm, and goes with say, 90MTr/mm, it would still be really small. Sub-100 range.

 

[Pokemaniac]

What else are we referring to? I’m only talking about sharing on the platform here.

On Switch, Nintendo has sometimes used Twitter as a replacement for where they previously used Miiverse for in game integration, and I think it's pretty safe to say now that that experiment is failing. General social media can be useful for boosting reach, but, as many companies are now finding out, the rug can be pulled out from under them quite quickly with little warning. It's best to use it without relying on it.