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

[SiG]

I'm really just wondering if the Ampere bandwidth budget for desktops applies to the T239.If there are so many variables then I have to wonder if we are being overly optimistic about the bandwidth needs of the T239.

There's also being too overly pessimistic.

 

[gabeplays]

AA needs to be enabled for DLSS to work. DLSS is a kind of AA.

Reminder that Nvidia literally has their own form of AA called DLAA or "Deep learning anti-aliasing". DLSS is "Deep Learning Super Sampling" which is a totally different thing and focuses on upscaling performance and resolution.

 

[SiG]

Reminder that Nvidia literally has their own form of AA called DLAA or "Deep learning anti-aliasing". DLSS is "Deep Learning Super Sampling" which is a totally different thing and focuses on upscaling performance and resolution.

Did you know that DLSS was originally concieved as an AA solution initially? I guess when they realized just how good an image quality they could achieve via AI reconstruction and pairing it up with reducing performance costs by upscaling from a lower resolution, they knew they were on to something.

DLAA basically ended up becoming the idea initially concieved (relying on machine learning to figure out anti-ailiasing without any AA "penalty" like what MSAA used to do).

 

[Machrider]

IMO the only thing we can do is to hope that Nintendo/Nvidia have chosen the right balance between cores/cache/bandwidth. They are the engineers and game developers. I honestly don't think anyone here can talk about this topic with confidence, especially when it comes to a custom SoC on a closed platform. All we can do is leave it in the hands of Nintendo/Nvidia.

The cores/bandwidth and all that is fine.

It's more about what Nintendo allows as clock speeds.

Because the chip itself is perfectly capable of running even PS5 tier titles, 1536 CUDA core Ampere based Nvidia GPU with DLSS? More bandwidth and RAM than a 2050 GPU that runs most PS5 games?

Yeah those aren't going to be the issues, the issue is particularly going to be the undocked mode, having a great chip doesn't matter much if they insist on tying its hands behind its back and only allowing like 5 watt power consumption.

That's why I'm all for full performance even undocked. External batteries are dirt cheap and ubiquitous these days, power outlets are available on airplanes, buses, trains, airports, hospitals, shopping malls, cafes, etc. etc. etc., the world has changed in terms of battery options, this isn't the Game Boy or even DS era any more where if you were out of power you were generally shit out of luck. It seems to me like the top USB charging port is a nod to that exact logic.

 

[Mr.Gamerson]

Quick somebody upload more videos of DLSS ultra performance from 360p to 1080p so we can see how well it holds up on different games. Upload to 4k if possible if using youtube so we can get access to the highest bitrate.

 

[Kise Ryota]

More bandwidth and RAM than a 2050 GPU

Actually the 2050 has 112GB/s (that's all for the GPU, while the T239 is expected to have 120GB/s shared between GPU and CPU (like it was said with more details in the last page)

Yeah those aren't going to be the issues, the issue is particularly going to be the undocked mode, having a great chip doesn't matter much if they insist on tying its hands behind its back and only allowing like 5 watt power consumption.

That's why I'm all for full performance even undocked.

What you are saying is like what was talked about the other, right? When people commented about what John (from DF) said.

I understand more hardcore players would like to go full performance undocked, especially those who play more in handheld than docked. What I don't like about this approach is that the resolution in docked mode would be the same as in handheld mode (1080p) because there wouldn't be any boost in clock.

But, like I said, I understand where you're coming from.

 

[NineTailSage]

To clarify, In this case you are confirming the post your responding today belief Nintendo will reduce cache in t239 as a potential cost saving measure?

Well that's interesting because it's an opposite statement lol, so we can't be sure yet that the T239 has the same amount of cache as the desktop Ampere right? Then we likewise can't determine the bandwidth requirements of the T239.

I think when it comes to Nintendo usually they do the opposite of whatever came before.
The Switch 1 was extremely bottlenecked, so history tells us Nintendo will go out of their way to make sure the successor won't be close to that situation again.

Just think after the N64, Nintendo gave us the GC which had very unique fast low latency 1t-SRAM and that system ran everything they threw at it... Fast forward to the WiiU coming off the Wii (yes it was a Frankenstein of a machine) but Nintendo put 32MB of eDRAM on the MCM which was super expensive to solve for slow RAM.

They have better options now and don't have to use expensive boutique memory, but I fully expect the CPU in Switch 2 to be a bottleneck before memory bandwidth ever does.

 

[wufuqin1999]

I think when it comes to Nintendo usually they do the opposite of whatever came before.
The Switch 1 was extremely bottlenecked, so history tells us Nintendo will go out of their way to make sure the successor won't be close to that situation again.

Just think after the N64, Nintendo gave us the GC which had very unique fast low latency 1t-SRAM and that system ran everything they threw at it... Fast forward to the WiiU coming off the Wii (yes it was a Frankenstein of a machine) but Nintendo put 32MB of eDRAM on the MCM which was super expensive to solve for slow RAM.

They have better options now and don't have to use expensive boutique memory, but I fully expect the CPU in Switch 2 to be a bottleneck before memory bandwidth ever does.

I don't know, I just don't know what evidence we have that the T239 requires the same bandwidth for 1tflops as a desktop Ampere. I would have assumed that bandwidth was determined by architecture if that Chinese netizen hadn't reminded me that cache affects bandwidth requirements. Given the cache cuts on the TX1 compared to the desktop Maxwell, it's possible that the same problem could be faced on the T239, and then the bandwidth required for 1tflops would be higher than 25gb/s.

 

[NineTailSage]

I don't know, I just don't know what evidence we have that the T239 requires the same bandwidth for 1tflops as a desktop Ampere. I would have assumed that bandwidth was determined by architecture if that Chinese netizen hadn't reminded me that cache affects bandwidth requirements. Given the cache cuts on the TX1 compared to the desktop Maxwell, it's possible that the same problem could be faced on the T239, and then the bandwidth required for 1tflops would be higher than 25gb/s.

Well look at the addition of the FDE, Nintendo and Nvidia have set T239 up to be streamlined in its functions.
Instead of relying on the CPU to perform decompression tasks, now this new hardware will handle those tasks.

We also just have to think about what resolutions this system will be targeting natively.
This device won't be looking for native 4k or anything like that and with DLSS it won't need to.

Edit: Also, we are still unaware of how much total system cache T239 will have access to...
They may have decided to go with a higher amount of SLC(system level cache), in order to let both the GPU and CPU have access to that memory vs high amounts of dedicated GPU cache.

 

[wufuqin1999]

Well look at the addition of the FDE, Nintendo and Nvidia have set T239 up to be streamlined in its functions.
Instead of relying on the CPU to perform decompression tasks, now this new hardware will handle those tasks.

We also just have to think about what resolutions this system will be targeting natively.
This device won't be looking for native 4k or anything like that and with DLSS it won't need to.

We're not talking about the same thing.

 

[NineTailSage]

We're not talking about the same thing.

I don't understand this statement
TX1 wasn't a custom Nintendo SoC, T239 is extremely customized for a gaming device.
I just think we won't have a full picture until we know the full details of everything in total...

 

[Concernt]

If I understand correctly, architecture is just a broad term, and the bandwidth requirements that specifically determine a particular architecture are still determined by the number of caches.

Cache is one element that contributes to efficient memory and bandwidth use, but does not necessarily itself contribute to bandwidth. Cache also doesn't, in and of itself, technically have a bandwidth requirement.

There also aren't a "number of caches" in a way that correlates to bandwidth - there are levels, and there are amounts. But you can have many layers, and a huge amount, but very low bandwidth - commonly, high performance mobile devices go the opposite direction, with less cache but at moderate or high bandwidth. The successor's memory bandwidth caps out at around 120GB/s regardless of the cache situation.

With it having relatively fast memory that's mounted on the same board as the die of the SOC, being so close, and with the GPU and CPU able to communicate directly (as they are the same chip), the need for cache on the successor is less than an equivalent PC targeting the same performance.

 

[Goodtwin]

I don't know, I just don't know what evidence we have that the T239 requires the same bandwidth for 1tflops as a desktop Ampere. I would have assumed that bandwidth was determined by architecture if that Chinese netizen hadn't reminded me that cache affects bandwidth requirements. Given the cache cuts on the TX1 compared to the desktop Maxwell, it's possible that the same problem could be faced on the T239, and then the bandwidth required for 1tflops would be higher than 25gb/s.

The truth is we do not know for sure, but we do have circumstantial evidence that support the idea that its not going to be an issue. If your looking for absolute definitive proof that there isnt a bandwidth bottleneck, nobody can provide that here. With that said, there is no evidence that T239 will be bottleneck in this way. Pointing to the Tegra X1, an off the shelf SOC that originally targeted the mobile market is very different from a custom SOC that targeted a gaming console from the very start.

 

[Deleted member 511]

here's a new video comparing UE5.4 with ray reconstruction on and off

It's going to be very interesting this generation comparing ray-tracing quality with consoles that can use ray reconstruction to consoles that rely on traditional, disparate denoisers. It isn't necessarily the case that a more powerful console will produce higher quality ray-tracing results, so I look forward to seeing comparisons like this in the future, especially with weaker hardware that can use ray reconstruction against more powerful hardware that can't. It's not something you can objectively quantify without a direct comparison and will likely depend on case-by-case.

 

[darthdiablo]

I don't know, I just don't know what evidence we have that the T239 requires the same bandwidth for 1tflops as a desktop Ampere. I would have assumed that bandwidth was determined by architecture if that Chinese netizen hadn't reminded me that cache affects bandwidth requirements. Given the cache cuts on the TX1 compared to the desktop Maxwell, it's possible that the same problem could be faced on the T239, and then the bandwidth required for 1tflops would be higher than 25gb/s.

You can't really compare things that way. TX1 was not a custom SoC, while T239 is. What happened on TX1 doesn't make it more likely same thing will happen with T239.

Also, I feel like NineTailSage is talking about the same thing. NineTailSage was pointing out examples where it's obviously a bit more nuanced than looking at what they did with TX1 (which again wasn't a custom SoC)

 

[albumin]

ROPS do something called pixel sampling. Most forms of AA heavily use pixel sampling.

DLSS does AA without pixel sampling. So using DLSS removes load on the ROPS.

In general, modern game engines are slowly using the ROPS less and less. You can expect Nintendo, who only has one platform to support, to move to that kind of an engine very quickly, and to not have a problem with limited ROPS.

T239's pixel sampling power is about halfway between the Xbox One and the PS4. The Xbox One was considered stingy with ROPS, and the PS4 was considered overly generous. It's probable that T239 is in a Goldilocks spot for last gen ports.

Where T239's pixel sampling looks a little stingy is relative to other RTX 30 cards. In fact, the stinginess makes it look like an intentional choice on Nintendo/Nvidia's part, as if they believed they didn't need the ROPS, either because Nintendo's own engines didn't need them, or because the chip space was better spent elsewhere.

That's why I don't believe it to be a problem for Nintendo games but possibly an issue for some lower budget ports that aren't putting in the money and time to really optimize for T239.

What other things tax ROPS heavily (and how heavily exactly; is AA usually the biggest culprit?) Also I'm wondering if FSR also uses pixel sampling (because if it also doesn't then a switch 2 with DLSS wouldn't really be lighter on ROPS than the other consoles are)

 

[NineTailSage]

It's going to be very interesting this generation comparing ray-tracing quality with consoles that can use ray reconstruction to consoles that rely on traditional, disparate denoisers. It isn't necessarily the case that a more powerful console will produce higher quality ray-tracing results, so I look forward to seeing comparisons like this in the future, especially with weaker hardware that can use ray reconstruction against more powerful hardware that can't. It's not something you can objectively quantify without a direct comparison and will likely depend on case-by-case.

I have to say, I wasn't expecting the difference in the details to be that big, but that was surprising!

 

[Mis4ha]

The cores/bandwidth and all that is fine.

It's more about what Nintendo allows as clock speeds.

Because the chip itself is perfectly capable of running even PS5 tier titles, 1536 CUDA core Ampere based Nvidia GPU with DLSS? More bandwidth and RAM than a 2050 GPU that runs most PS5 games?

Yeah those aren't going to be the issues, the issue is particularly going to be the undocked mode, having a great chip doesn't matter much if they insist on tying its hands behind its back and only allowing like 5 watt power consumption.

That's why I'm all for full performance even undocked. External batteries are dirt cheap and ubiquitous these days, power outlets are available on airplanes, buses, trains, airports, hospitals, shopping malls, cafes, etc. etc. etc., the world has changed in terms of battery options, this isn't the Game Boy or even DS era any more where if you were out of power you were generally shit out of luck. It seems to me like the top USB charging port is a nod to that exact logic.

The world isn’t always on battery, the world is always plugged in.

 

[oldpuck]

Is Switch 2 an Xbox?

I'm pretty sure Microsoft would love it to be, if they could convince Nintendo to let them get GamePass on it, they would.

It's going to be very interesting this generation comparing ray-tracing quality with consoles that can use ray reconstruction to consoles that rely on traditional, disparate denoisers. It isn't necessarily the case that a more powerful console will produce higher quality ray-tracing results, so I look forward to seeing comparisons like this in the future, especially with weaker hardware that can use ray reconstruction against more powerful hardware that can't. It's not something you can objectively quantify without a direct comparison and will likely depend on case-by-case.

Looking at the frame rate, RR seems to have a ~2.5ms cost in these examples, on a 4080. It'll be interesting to see how much RR we actually see on Switch 2. It'll also be interesting to see FSR 4's solution, since they recently published some work on unified neural denoising, and whether or not any of that makes it to existing consoles.

 

[oldpuck]

What other things tax ROPS heavily (and how heavily exactly; is AA usually the biggest culprit?) Also I'm wondering if FSR also uses pixel sampling (because if it also doesn't then a switch 2 with DLSS wouldn't really be lighter on ROPS than the other consoles are)

The short answer is "everything uses ROPS". A medium sized answer involves getting deep into the woods of how rendering works (and pushing my own knowledge to the limit).

Imagine a grassy lawn, Actually, don't bother imagining it, here is a picture for you.

Okay, what resolution is the lawn? Not the resolution of my picture of the lawn, but the resolution of the lawn itself, in real life. Deep question, but roll with it for a second. From our point of view, the resolution is infinity. No matter how close you get to a piece of grass, you never see the pixels. In fact, pixels don't exist, the real world isn't made of pixels.

Deep inside your GPU, it's actually the same thing. A 3D scene isn't made of pixels, it's made of 3D objects, which themselves are made of triangles, and just like the real world, triangles have infinite resolution. You can zoom a triangle in forever, and always get a smooth line and sharp corners.

But your screen is made of pixels. We need a way to convert from the infinite resolution to finite resolution. From that perfectly smooth triangle, to a pixelated one.

How do we do that? Well, let's go back to our lawn. Imagine that lawn* with a fence in front of it.

Okay, now what? Well, we've broken the infinite resolution lawn up into a pixel grid. Now we just need to determine the color of each pixel. How? By sampling it. Essentially imagine taking a laser, and shooting right in the middle of each diamond on the grid, and whatever color you hit, that's the color of the whole pixel.

You can see right now how much detail you'd lose that way. Just from this picture alone, almost every pixel would be green, except for a few brown ones. All the details of the individual blades of grass would go away, a lot of the patchiness of the lawn would be replaced by a single boring green. And the few brown patches we got would be super blocky.

What if we wanted to smooth out the edges of things - anti alias them? Well, we could shoot multiple lasers just at the edges of objects we wanted to smooth out - multi sampling. What if we wanted to do more than just show green? We could blend colors together.

This whole process is called rasterization. ROPs stands for Raster OPerators and this is their job. And when rendering a single frame of a video game, modern game engines don't just rasterize one image, they rasterize dozens. An image for shadows, an image for highlights, an image for the background, an image for the foreground, an image for all the metal objects in a scene, and image for all the cloth objects, etc etc.

And these images are combined, and the ROPs do that as well, blending all these different images (called "buffers") together to make the final frame you see on screen. That's the way that rendering has worked in GPUs since basically the very beginning.

Except... programmers are starting to figure they can do this process better than the hardware designers. Or at the very least, they can do better than the one-size-fits-all that the ROPs provide, and build a more efficient solution that is customized just for their game. Instead they're using the shader cores to do some or all of these operations, reducing the lean on ROPs.

DLSS is a specialized example of that. Instead of drawing a 4k image using ROPs, you draw a 1080p image (with or without ROPs), and then upscale it using the tensor cores, totally separate from the old ROP pipeline. This moves huge chunks of work away from the ROPs

Does all that make sense?

* Or a very similar one, I'm kinda limited by Google image search here.

 

[albumin]

The short answer is "everything uses ROPS". A medium sized answer involves getting deep into the woods of how rendering works (and pushing my own knowledge to the limit).

Imagine a grassy lawn, Actually, don't bother imagining it, here is a picture for you.

Okay, what resolution is the lawn? Not the resolution of my picture of the lawn, but the resolution of the lawn itself, in real life. Deep question, but roll with it for a second. From our point of view, the resolution is infinity. No matter how close you get to a piece of grass, you never see the pixels. In fact, pixels don't exist, the real world isn't made of pixels.

Deep inside your GPU, it's actually the same thing. A 3D scene isn't made of pixels, it's made of 3D objects, which themselves are made of triangles, and just like the real world, triangles have infinite resolution. You can zoom a triangle in forever, and always get a smooth line and sharp corners.

But your screen is made of pixels. We need a way to convert from the infinite resolution to finite resolution. From that perfectly smooth triangle, to a pixelated one.

How do we do that? Well, let's go back to our lawn. Imagine that lawn* with a fence in front of it.

Okay, now what? Well, we've broken the infinite resolution lawn up into a pixel grid. Now we just need to determine the color of each pixel. How? By sampling it. Essentially imagine taking a laser, and shooting right in the middle of each diamond on the grid, and whatever color you hit, that's the color of the whole pixel.

You can see right now how much detail you'd lose that way. Just from this picture alone, almost every pixel would be green, except for a few brown ones. All the details of the individual blades of grass would go away, a lot of the patchiness of the lawn would be replaced by a single boring green. And the few brown patches we got would be super blocky.

What if we wanted to smooth out the edges of things - anti alias them? Well, we could shoot multiple lasers just at the edges of objects we wanted to smooth out - multi sampling. What if we wanted to do more than just show green? We could blend colors together.

This whole process is called rasterization. ROPs stands for Raster OPerators and this is their job. And when rendering a single frame of a video game, modern game engines don't just rasterize one image, they rasterize dozens. An image for shadows, an image for highlights, an image for the background, an image for the foreground, an image for all the metal objects in a scene, and image for all the cloth objects, etc etc.

And these images are combined, and the ROPs do that as well, blending all these different images (called "buffers") together to make the final frame you see on screen. That's the way that rendering has worked in GPUs since basically the very beginning.

Except... programmers are starting to figure they can do this process better than the hardware designers. Or at the very least, they can do better than the one-size-fits-all that the ROPs provide, and build a more efficient solution that is customized just for their game. Instead they're using the shader cores to do some or all of these operations, reducing the lean on ROPs.

DLSS is a specialized example of that. Instead of drawing a 4k image using ROPs, you draw a 1080p image (with or without ROPs), and then upscale it using the tensor cores, totally separate from the old ROP pipeline. This moves huge chunks of work away from the ROPs

Does all that make sense?

* Or a very similar one, I'm kinda limited by Google image search here.

It does make sense, thanks a lot for the explanation I appreciate it
going off of this, FSR would also take some load off of the ROPS correct, since it also renders at a lower resolution, and provides anti-aliasing to the final picture

 

[TomNookYankees]

I only just now noticed that nintendo's "Switch Successor will play Switch software" tweet was worded such that cart compatibility is still up in the air.

 

[darthdiablo]

I only just now noticed that nintendo's "Switch Successor will play Switch software" tweet was worded such that cart compatibility is still up in the air.

Nintendo refers to both digital and physical games as “software”

 

[Sin & Punishment]

I only just now noticed that nintendo's "Switch Successor will play Switch software" tweet was worded such that cart compatibility is still up in the air.

No, it's worded to be inclusive. (physical + digital)

If they said "Switch Successor will play Switch cartridges" then you'd have the complete opposite problem

 

[omniryu]

Switch 2 will be more like the Gameboy Advance taking over from Gameboy/Gameboy Color line. Like the Gameboy Advance, it follows a console that has had a much longer life than expected, at long last providing a significant bump to performance but still a budget-friendly gaming device (with full BC!) in a market with higher spec options. I don't think Switch 2 will be the Lynx of the market. Even the OG Steamdeck will likely cost more than a Switch 2 and we'll likely see even pricier, high spec android and windows handhelds released in 2025.

Except the Switch 2 might have a longer lifespan. The reason why the GBA had shorter lifespan is because the tech was there yet. Nintendo had an influence on ARM. The Thumb instruction set was because using the regular arm instruction would add more to the storage of their games.. Did You Know Gaming did a video on Project Atlantis, the successor to the Gameboy and the processor was basically at fault.

Now the reason why the Switch 2 will last longer is because development takes longer. Games last longer, the "forever games" (see fortnite). We don't have 5 year consoles anymore or else the PS5 Pro would have been a PS6.

I am the only one who think this, but I think we are near peak visual satisfaction. I have a hard time seeing Switch 4 and PS7. I have a hard time seeing Switch 3 but who knows.

 

[skywake]

GBC is just the GB with everything doubled

It would fit well with the Switch Pro we didn't get but Switch 2 is completely next gen.

I would point out that I did very deliberately qualify my post by saying "in some ways". Yes Switch 2 is a significant performance jump and a "next gen" which on paper makes it more like the GB->GBA than GB->GBC. But in terms of game compatibility and the legacy software experience GB->GBC is far closer to what we're going to end up seeing

And on the OLED side sure you can argue that it was just a mid gen revision like the DSi, New 3DS and GBC before it. But the OLED only brought a better screen, a more robust case/stand and improved battery life. So in that sense it's more like the DSLite, 3DSXL or GB Pocket

As someone else said, these comparisons are a bit anachronistic and so don't line up. This transition is a modern hardware transition and it's unlike anything Nintendo has done before in many ways. The better comparisons are with XB Series and PS5. But I do think it's worth highlighting how, from a consumer perspective, this will feel more like a hardware revision than a new console when they're scrolling through the eShop on day1. And that's a good thing

 

[karmitt]

I only just now noticed that nintendo's "Switch Successor will play Switch software" tweet was worded such that cart compatibility is still up in the air.

We’re trying to find negatives where there are none. I just don’t see them hiding the fine print saying [**digital only].

If there are negative caveats to it, I expect them to be small. Like Labo titles won’t work because of the tablet size, or how you need the original Switch Joycons for Ring Fit.

 

[Concernt]

We’re trying to find negatives where there are none. I just don’t see them hiding the fine print saying [**digital only].

If there are negative caveats to it, I expect them to be small. Like Labo titles won’t work because of the tablet size, or how you need the original Switch Joycons for Ring Fit.

Labo would be compatible in the same way its compatible with Lite and OLED - imperfect but playable.

Plus, it's Labo. You can customise it to fit if you have to!

Ring Fit is a big one though, I wonder what they'll do about it given it's the big evergreen without native compatibility. I have to imagine we get a Ring Fit game early. Expanded Update a-la Wii Fit Plus/U, or sequel? Maybe a question for the other thread, but I'd imagine the Ring-Con gets redesigned significantly to move the controller around.

 

[SiG]

I don't know, I just don't know what evidence we have that the T239 requires the same bandwidth for 1tflops as a desktop Ampere. I would have assumed that bandwidth was determined by architecture if that Chinese netizen hadn't reminded me that cache affects bandwidth requirements. Given the cache cuts on the TX1 compared to the desktop Maxwell, it's possible that the same problem could be faced on the T239, and then the bandwidth required for 1tflops would be higher than 25gb/s.

The TX1 did not see any cuts in its cache when compared with desktop Maxwell. You have your information wrong.

 

[SMD]

i wonder if something like ring fit gets a sequel that accommodates all joycons, will it really need the power of the switch 2? that and switch sports 2 are prime candidates for crossgen imo

 

[gabeplays]

Now the reason why the Switch 2 will last longer is because development takes longer. Games last longer, the "forever games" (see fortnite). We don't have 5 year consoles anymore or else the PS5 Pro would have been a PS6.

While those are certainly valid reasons, that isn't the main core reason why the Switch continues to be supported for so long: it keeps making money.

This is made crystal clear in the recent Investor's Meeting. The Switch era has been a renaissance for Nintendo, especially as it's helped fuel their efforts to transform into a full-blown entertainment company with museums, theme parks, movies, and alarm clocks. They themselves outline that the sheer success of the Switch has helped them to push away from the traditional "video game business cycle" and for good reasons: what would be the point of pushing out a successor or Switch Pro out the door, like how Sony is now releasing their PS5 Pro, when the OG Switch continues to print money like hotcakes? when you can instead continue to give new and old players more value for their money by continuing to pump out new and exciting titles on the regular which they've absolutely nailed?

This is a primary reason why backwards compatibility is a thing. Because they want to continue their momentum and keep adding more value to consumers by allowing them to continue to bring forth their titles. It means players won't have to immediately purchase the successor and can hold onto their Switch for even longer periods of time which they want. As long as people are buying, they'll continue to support the OG Switch with new titles and cross-gen titles. Adding value is the most important thing for Nintendo. They want more people to buy consoles, to stay longer and play longer. It's why Nintendo Online Accounts are a thing. It's why merging their handheld and console divisions was the smartest dang thing they've ever done because they can put all their efforts into one single console and flood us up to our eyeballs with games. As long as people are buying in droves, they don't really care much for what Sony or Microsoft are doing or what the "traditional dedicated video game platform business cycle" is supposed to be. As they themselves say: "they are building a new foundation for the future" where they chart their own course and decide for themselves how long things should last.

Nintendo firmly wants to march to the beat of their own drum. They aren't blind to change or what the rest of the industry is doing, or a Switch 2 wouldn't be around the corner, but it's undeniable that a reason they've taken their sweet time is because what's the rush? I understand a lot of hardcore gamers are anxious to move onto the next shiny thing. It's what we've been used to for our whole lives. Consoles are generally only supposed to last five years or so: but is anybody really complaining that we've had even more games to play this year and we know Xenoblade, Metroid Prime 4, Pokemon, and other games are due to arrive next year?

Here in late 2024, it's starting to be a different story. Sales are finally starting to slow down to pre-pandemic numbers (which is still very good btw), talk of hardware limitations has become louder such as with Echoes of Wisdom frame-rates; now is the time to start putting out a successor which they are going too. But I don't blame Nintendo for not being in a rush because, in a very real way, they've left the console wars a long time ago and are only competing against themselves. They are literally as successful as they've ever been as a company and the Switch lasting until it's 8th year is picture proof of that. As much as I loath some of their decisions such as being so viscerally against fan content, they've done so many things right in the Switch era and I'm excited to see where they go with the Switch successor.

 

[fwd-bwd]

I was asked offline by a few folks to clarify certain points in my post about Nintendo’s 7 billion yen of work-in-progress inventory (about 44.8 million USD). Since it might be useful to others as well, I’m sharing some answers below. Pardon the random capitalization for readability.

• After a part or component from a supplier is in Nintendo’s possession, it's counted toward the Raw Materials inventory.
  • It matters not if the material is in a warehouse, on a cargo ship, or even sitting on a factory floor ready to be used, it is still in the Raw Materials category.
• When the material is transformed into part of a product (and the product isn’t completely finished), it is now counted toward the Work-in-Progress inventory.
  • For example, after Nintendo receives a T239 SoC from Nvidia, it is included in Raw Materials. As soon as the chip is mounted on a PCB, it becomes WIP.
• Knowing the above, the 7 billion yen of WIP inventory is probably the smoking gun of Switch 2 mass production as of 2024-09-30.
  • Not only it’s the highest WIP in the Switch era, but the amount is close to 2x of the previous high (when Switch 1 was selling like hotcakes during the pandemic).
  • Since Alarmo is priced at $99 and its inventory seems low, it likely didn’t have much impact on the WIP number.
• Note that from the number we can only say that some pre-assembly was underway, but it isn’t certain that the final assembly of a complete retail product had begun in September.
  • As LiC pointed out, September shipping data suggests that Hosiden Vietnam (sub)contracted some PCB SMT work to Guangdong Huazhuang Vietnam and Joy-Con strap work to Matsuoka Vietnam.
  • It’s quite possible that similar pre-assembly also had started in China. I wouldn’t be surprised if it was earlier and at a larger scale than what the Vietnamese data indicated.
• Previously I posted a few Nintendo suppliers’ financial reports (Genesys Logic, Weltrend, PixArt, and MegaChips) that offered hints of suppler mass production in Q3. The schedule lines up with pre-assembly in September.
• Just to be clear, I’m not saying that the final assembly had not commenced in September, only that we have no data to confirm it.
• A side note: Materials used in the construction of Nintendo Development Building 2 should fall under the capital expenditures, not in the inventory.

Edit: typo

 

[fwd-bwd]

Did my post about the start of pre-assembly and possibly final assembly remind you of what that self-proclaimed factory worker said back in September?

 

[LoonTwink]

I was asked offline by a few folks to clarify certain points in my post about Nintendo’s 7 billion yen of work-in-progress inventory (about 44.8 million USD). Since it might be useful to others as well, I’m sharing some answers below. Pardon the random capitalization for readability.

• After a part or component from a supplier is in Nintendo’s possession, it's counted toward the Raw Materials inventory.
  • It matters not if the material is in a warehouse, on a cargo ship, or even sitting on a factory floor ready to be used, it is still in the Raw Materials category.
• When the material is transformedinto part of a product (and the product isn’t completely finished), it is now counted toward the Work-in-Progress inventory.
  • For example, after Nintendo receives a T239 SoC from Nvidia, it is included in Raw Materials. As soon as the chip is mounted on a PCB, it becomes WIP.
• Knowing the above, the 7 billion yen of WIP inventory is probably the smoking gun of Switch 2 mass productionas of 2024-09-30.
  • Not only it’s the highest WIP in the Switch era, but the amount is close to 2x of the previous high (when Switch 1 was selling like hotcakes during the pandemic).
  • Since Alarmo is priced at $99 and its inventory seems low, it likely didn’t have much impact on the WIP number.
• Note that from the number we can only say that some pre-assembly was underway, but it isn’t certain that the final assemblyof a complete retail product had begun in September.
  • As LiC pointed out, September shipping data suggests that Hosiden Vietnam (sub)contracted some PCB SMT work to Guangdong Huazhuang Vietnam and Joy-Con strap work to Matsuoka Vietnam.
  • It’s quite possible that similar pre-assembly also had started in China. I wouldn’t be surprised if it was earlier and at a larger scale than what the Vietnamese data indicated.
• Previously I posted a few Nintendo suppliers’ financial reports (Genesys Logic, Weltrend, PixArt, and MegaChips) that offered hints of suppler mass production in Q3. The schedule lines up with pre-assembly in September.
• Just to be clear, I’m not saying that the final assembly had not commenced in September, only that we have no data to confirm it.
• A side note: Materials used in the construction of Nintendo Development Building 2 should fall under the capital expenditures, not in the inventory.

Edit: typo

awesome post, ty! how is WIP inventory valued? Is it just the sum of the raw materials, is it some estimation of the "value of work done so far", or maybe retail/wholesale value minus the remaining parts required? Not that this matters an incredible amount, given that 7 billion yen/$45mm implies a whole lot of something, but I was wondering if we could infer a number of units included in this WIP inventory. Depending on the accounting it could swing by a factor of roughly 2 I would guess, and then depending on what exactly was done it could have another factor of uncertainty.

 

[Nigel Kidd]

The comparison with the GameBoy is obvious but I'm not sure the OLED and Switch 2 line up that well with GBC and GBA. In some ways the OLED is more like the GB Pocket and in some ways the Switch 2 will be more like the GBC

What I mean by that is that the Pocket was basically just a more modern and power efficient screen. Which, combined with improved battery tech, meant they could drop down to AAA batteries. It was a cut-down model but it was also generally an improved piece of hardware. More or less what the OLED also did. Yes the GBC was also a screen improvement but it also had more RAM and a higher CPU clock. The OLED didn't really bring any improved specs, at least officially

And for Switch 2 we're going to get full library compatibility and presumably a fairly seamless experience in that regard. That wasn't true when playing GB games on the GBA because the GBA had an entirely different architecture and an unfriendly for GB(C) screen resolution. Backwards compatibility only existed because they physically included the GBC CPU, on most models. Sure the jump in power is more analogous to the GBA but I think the end-user experience will be closer to the GBC

Your analogy is also quite good. It's probably the better one if we're just talking Switch + Switch OLED which do map better on to GB/GB Pocket or maybe even DS/DSi. My analogy is only meant to be quite general. The fact that it's difficult to say whether the leap is akin to GB>GBC or GBC>GBA reflects the fact that generational leaps are very different things today. GBC was very much an update to the existing paradigm. We saw it as GB but in colour. GBA was a generational leap. We saw it as SNES in your pocket which seemed amazing. It remains to be seen I suppose which category Switch 2 will fall into. Will it feel like a generational leap or a Switch Max Pro?

Just want to also add that I don't think the GBA screen was 'unfriendly' for GB/C screen resolution at all. It still displayed pixel perfect & in the same manner. I would argue it was far friendlier to GB/C than the Switch screen is to N64 since it doesn't have the CRT scanlines + mushiness which you need to blend the polygons and textures into a pleasing visual. But I am opening a whole other can of worms there, aren't I?

 

[fwd-bwd]

how is WIP inventory valued? Is it just the sum of the raw materials, is it some estimation of the "value of work done so far", or maybe retail/wholesale value minus the remaining parts required?

The WIP inventory (perhaps a bit counterintuitively) also includes the labor and overhead costs for the manufacturing. Since Nintendo outsourced this to Foxconn and Hosiden, instead of accounting for the labor and overhead costs directly, the WIP inventory should include the assembly fees paid to the assemblers (aka contract manufacturing costs).

 

[LoonTwink]

The WIP inventory (perhaps a bit counterintuitively) also includes the labor and overhead costs for the manufacturing. Since Nintendo outsourced this to Foxconn and Hosiden, instead of accounting for the labor and overhead costs directly, the WIP inventory should include the assembly fees paid to the assemblers (aka contract manufacturing costs).

So all the assembly fees up front? How much per unit would that roughly be? I would guess in the ballpark of 5-15% but that's spitballing

 

[omniryu]

While those are certainly valid reasons, that isn't the main core reason why the Switch continues to be supported for so long: it keeps making money.

I was referring to the second switch. The person I responded to stated that the successor will make less money because like the Gameboy Advance,
The Switch 2 will be out longer.

I guess what you put on can be an addendum to what I posted.

 

[Deleted member 511]

I have to say, I wasn't expecting the difference in the details to be that big, but that was surprising!

I think a lot of people don't expect the differences to be very noticeable, but the thing about AI denoisers is that they're not just a means to get rid of the performance cost of running multiple denoisers at the same time but to understand the context of the inputs so that the denoising is more intelligently applied, reducing the occurrence of artifacts. That sort of holistic approach seems very promising to me.

Looking at the frame rate, RR seems to have a ~2.5ms cost in these examples, on a 4080. It'll be interesting to see how much RR we actually see on Switch 2. It'll also be interesting to see FSR 4's solution, since they recently published some work on unified neural denoising, and whether or not any of that makes it to existing consoles.

Definitely interested to see how devs tackle this as well. The great thing about the closed environment is that you can optimize your engine (or a version of it) around that hardware to reduce the overhead for both DLSS and RR (or equivalents). The question remains which devs are actually going to do that. Such targeted optimization needs ample budget, time, and resources that not every dev will have.

 

[BDGAME]

It does make sense, thanks a lot for the explanation I appreciate it
going off of this, FSR would also take some load off of the ROPS correct, since it also renders at a lower resolution, and provides anti-aliasing to the final picture

FSR don't provide Anti Aliasing.

 

[oracion]

What is the difference between DLSS 3.5 and 3.7?

 

[ILikeFeet]

What is the difference between DLSS 3.5 and 3.7?

I'm assuming you're confusing the marketing with the version numbering. marketing "3.5" is the addition of Ray Reconstruction to the suite (which includes Super Resolution and Frame Generation). versioning is just an updated Super Resolution (and ray reconstruction). Frame Generation is its own DLL if I remember correctly

 

[oracion]

I'm assuming you're confusing the marketing with the version numbering. marketing "3.5" is the addition of Ray Reconstruction to the suite (which includes Super Resolution and Frame Generation). versioning is just an updated Super Resolution (and ray reconstruction). Frame Generation is its own DLL if I remember correctly

I see! Thanks.

 

[omniryu]

Does all that make sense?

Unrelated question. Texture mapping Units. Help me out here, do they define the GPU max Texture resolution or no?

 

[oldpuck]

Unrelated question. Texture mapping Units. Help me out here, do they define the GPU max Texture resolution or no?

I believe modern TMUs can handle any resolution you can fit into memory, so no.

 

[kasparov77]

Labo would be compatible in the same way its compatible with Lite and OLED - imperfect but playable.

Plus, it's Labo. You can customise it to fit if you have to!

Ring Fit is a big one though, I wonder what they'll do about it given it's the big evergreen without native compatibility. I have to imagine we get a Ring Fit game early. Expanded Update a-la Wii Fit Plus/U, or sequel? Maybe a question for the other thread, but I'd imagine the Ring-Con gets redesigned significantly to move the controller around.

do switch 2 joycons not have an ir camera/sensor at the bottom for ringfit heart rate monitoring?

also while we're at it: is there any indication in customs data of a camera at all?

 

[kasparov77]

I would point out that I did very deliberately qualify my post by saying "in some ways". Yes Switch 2 is a significant performance jump and a "next gen" which on paper makes it more like the GB->GBA than GB->GBC. But in terms of game compatibility and the legacy software experience GB->GBC is far closer to what we're going to end up seeing

And on the OLED side sure you can argue that it was just a mid gen revision like the DSi, New 3DS and GBC before it. But the OLED only brought a better screen, a more robust case/stand and improved battery life. So in that sense it's more like the DSLite, 3DSXL or GB Pocket

As someone else said, these comparisons are a bit anachronistic and so don't line up. This transition is a modern hardware transition and it's unlike anything Nintendo has done before in many ways. The better comparisons are with XB Series and PS5. But I do think it's worth highlighting how, from a consumer perspective, this will feel more like a hardware revision than a new console when they're scrolling through the eShop on day1. And that's a good thing

best comparison i can think of is DSi -> 3DS

(if we ignore the massive storage and RAM gap)

 

[albumin]

FSR don't provide Anti Aliasing.

according to AMD, 'FSR 2 uses cutting-edge temporal algorithms to reconstruct fine geometric and texture detail, producing anti-aliased output from aliased input' 'Includes high-quality anti-aliasing. FSR 2 replaces any TAA within the game frame.'
I was kinda wondering here, since the other consoles use FSR I thought that DLSS wouldn't provide an advantage over that specifically in terms of usage of ROPS. We'll just have to see how heavy current and future games are on them i suppose

 

[Hartmann]

Here’s an interesting comment I found on Macronix's finance board.
Although it seems to contain a lot of speculation, I thought I’d share it here since the content is intriguing.

ä¸è½ä¸æç½!(11.14ä¿® æéæ¨èæ¼²è·ç¡é)

[ä¸è½ä¸æç½!(11.14ä¿® æéæ¨èæ¼²è·ç¡é)]æºå® æ¯ä¸çå¯ä¸ 3D Norãå¨èºå¯ä¸ 3D Nandçç§æå¬å¸ï¼ç¨ eps3.5å/å¹´ å»æç ç¼!ç¨å®¶ç ç¼ç SG VCæ¶æ§å¨ç¸åå ç層æ¸ä¸çå²åå¯åº¦æ¯å¶ä» GAA VC 3D Nandç 2~

www.cmoney.tw

Macronix is the only company globally to produce 3D NOR and the only one in Taiwan to produce 3D NAND, investing annual earnings of NT$3.5 per share into R&D! Its uniquely developed SG VC architecture achieves 2-3 times the storage density of other GAA VC 3D NAND at the same layer count.
On 2024/09/05, the United States restricted the export of 3D NAND technology (GAAFET - Gate-All-Around Field-Effect Transistors) to China.
On 2024/09/08, following this, Japan also restricted the export of GAAFET technology used in the design and manufacturing of integrated circuits to China.
On 2024/10/11, major 3D NAND manufacturers, Kioxia, Micron, Samsung, and SK Hynix, lowered production output, shifting some lines to focus on HBM. Samsung might convert its P4 line, while SK Hynix plans to convert its Cheongju M14, M15X, and M16 facilities.
On 2024/10/31, the EU raised import tariffs on Chinese electric vehicles from 10% to 45.3%; each electric vehicle contains 7,000 chips.
On 2024/11/08, Kioxia projected that NAND demand will triple between 2024 and 2029, with a compound annual growth rate of 5.43%. Starting in 2025, the United States will double semiconductor tariffs on China to 50%, and the strict assessment will shift from the final assembly location back to the front-end manufacturing and photomask origins. Simultaneously, the Netherlands will cease ASML maintenance for China.

Around 99% of 3D NAND production relies on GAAFET design software and inspection equipment, which disadvantages Chinese companies like Samsung, Hynix, and YMTC.
1) Samsung's two factories in Xi'an have a combined capacity of 250,000 wafers per month, limited to 128 layers or less, representing about 10% of global NAND flash capacity.
2) SK Hynix's (formerly Intel's) Dalian facility, also limited to 128 layers or less, represents about 6.8% of global NAND flash.
3) YMTC, constrained by the last round of restrictions from 232 to 162 layers, currently represents about 8% of global NAND flash. Regardless of the number of layers, the entire 3D NAND supply chain will be severed, with all Chinese GAAFET-dependent factories restricted and tariffed—double the cost for those who don’t relocate.

^O.O^ ^O.O^ Attention! Attention! The first revenue “bomb” is here! ^O.O^ ^O.O^
On 2021/04/12, Nikkei reported Nintendo's delay with a new console, quoting Furukawa: “To attract players to choose to game in their limited time, games must be fun,” while noting technical and cost limitations that had hindered certain ideas from being realized.
On 2024/07/30, Mr. Lu indicated that 3D NAND with 192 layers would ship by the end of 2024 or early 2025 (to whom? See below). Macronix stated that current 3D NAND capacity cannot meet market demand and is expected to improve by the latter half of next year.
On 2024/05/07, Furukawa announced that Switch 2 will be unveiled before 2025/03/31, with Macronix potentially winning a contract for 3D NAND. On 2024/11/06, Furukawa confirmed that Switch 2 will support backward compatibility and offer NSO services.
It's 99% certain that Macronix's 3D NOR will be used, with 3D NAND following behind it in a configuration that doesn’t allow for the substitution with other brands. The 3D NOR kicks open the door, and the 3D NAND follows, making the combo irresistible—a “Memory System” as Mr. Wu calls it.

Before Nintendo pulls stock, they’ll build inventory to an extremely high level, potentially distorting the financial report as Nintendo holds zero stock, with the entire inventory pushed onto Macronix. The new Switch 2 order for 3D NAND and ROM has increased from 32GB to 256GB, multiplying the original Switch’s bit count by 8-16 times. On 2024/07/01, Furukawa said there’s been no change in their view since the supply issues around Animal Crossing, implying a stockpile far exceeding 28.5 million units. The Q3 report also leaked that Switch 2 production started at least a month ago, with an estimated five-month lead time before release and the earliest deliveries expected by January!

They’ll need to account for inventory loss impacting gross margins while also ensuring a five-month stock turnaround, concentrated before the announcement. Issuing NT$3 billion in corporate bonds to repay the bank? It’s likely an indicator of substantial stock buildup to support pre-launch logistics.

China’s memory module industry relies on Chinese-produced NAND and SSD components. With China struggling with 3D production, Macronix will uniquely benefit as a “Fab-Lite.” In 2023, global revenue in this field reached $22 billion, and Macronix, in a supply-chain restructuring, has seized a huge opportunity for growth!

DRAM and HBM come with lifetime warranties, while NAND typically has a five-year lifespan, making it a consumable with favorable long-term revenue trends. This isn’t advice to buy! But if you buy, rest easy—sleep well and hold it tight! Rumors about insiders are meant to drive you to sell at a loss.

Macronix is soaring! Macronix is soaring! Macronix is soaring! I’m saying it three times for all those investors burned by insiders.

Extra Info:
Macronix’s two closest partners—Nintendo and IBM—are both major 3D NAND users. In the past, they turned to Samsung and Micron for supply, even after Macronix developed capabilities, but changes have come. Samsung’s manufacturing and Micron’s testing are now restricted, and with a 50% tariff, refusing to switch suppliers is costly. Furthermore, IBM has agreed to collaborate on 300+ layer 3D NAND controllers, and if they don’t turn to Macronix, they’ll just gain another competitor.

With these orders secured, Macronix’s revenue will skyrocket to a level that will leave you questioning reality. This isn’t advice to buy, but if you already own shares, hold onto them and sleep easy!

 

[MP!]

The short answer is "everything uses ROPS". A medium sized answer involves getting deep into the woods of how rendering works (and pushing my own knowledge to the limit).

Imagine a grassy lawn, Actually, don't bother imagining it, here is a picture for you.

Okay, what resolution is the lawn? Not the resolution of my picture of the lawn, but the resolution of the lawn itself, in real life. Deep question, but roll with it for a second. From our point of view, the resolution is infinity. No matter how close you get to a piece of grass, you never see the pixels. In fact, pixels don't exist, the real world isn't made of pixels.

Deep inside your GPU, it's actually the same thing. A 3D scene isn't made of pixels, it's made of 3D objects, which themselves are made of triangles, and just like the real world, triangles have infinite resolution. You can zoom a triangle in forever, and always get a smooth line and sharp corners.

But your screen is made of pixels. We need a way to convert from the infinite resolution to finite resolution. From that perfectly smooth triangle, to a pixelated one.

How do we do that? Well, let's go back to our lawn. Imagine that lawn* with a fence in front of it.

Okay, now what? Well, we've broken the infinite resolution lawn up into a pixel grid. Now we just need to determine the color of each pixel. How? By sampling it. Essentially imagine taking a laser, and shooting right in the middle of each diamond on the grid, and whatever color you hit, that's the color of the whole pixel.

You can see right now how much detail you'd lose that way. Just from this picture alone, almost every pixel would be green, except for a few brown ones. All the details of the individual blades of grass would go away, a lot of the patchiness of the lawn would be replaced by a single boring green. And the few brown patches we got would be super blocky.

What if we wanted to smooth out the edges of things - anti alias them? Well, we could shoot multiple lasers just at the edges of objects we wanted to smooth out - multi sampling. What if we wanted to do more than just show green? We could blend colors together.

This whole process is called rasterization. ROPs stands for Raster OPerators and this is their job. And when rendering a single frame of a video game, modern game engines don't just rasterize one image, they rasterize dozens. An image for shadows, an image for highlights, an image for the background, an image for the foreground, an image for all the metal objects in a scene, and image for all the cloth objects, etc etc.

And these images are combined, and the ROPs do that as well, blending all these different images (called "buffers") together to make the final frame you see on screen. That's the way that rendering has worked in GPUs since basically the very beginning.

Except... programmers are starting to figure they can do this process better than the hardware designers. Or at the very least, they can do better than the one-size-fits-all that the ROPs provide, and build a more efficient solution that is customized just for their game. Instead they're using the shader cores to do some or all of these operations, reducing the lean on ROPs.

DLSS is a specialized example of that. Instead of drawing a 4k image using ROPs, you draw a 1080p image (with or without ROPs), and then upscale it using the tensor cores, totally separate from the old ROP pipeline. This moves huge chunks of work away from the ROPs

Does all that make sense?

* Or a very similar one, I'm kinda limited by Google image search here.

Didn’t realize every time I look out my screen door I’m rasterizing my backyard
Great explanation