Topic: The Nintendo Switch Rumor and Speculation Thread

@Discostew True. That's why I'm concerned Nintendo will stick with the Samsung 8N node, and thusly performance will have a lower ceiling than we're hoping.

Theoretically, how much time do you think Nintendo would need to switch to TSMC in order to start mass production in late 2024? Would they need to make the switch as early as 2022?

@Ulysses T239 was taped out in August 2022, so the node decision was probably made early 2021, if not earlier. I mostly agree with @Discostew, I don’t think they changed the node after they committed the design. 8nm was possibly a cheap option offered to Nintendo back in 2019, or it would have been ruled out at the early design phase once they were able to simulate it’s power and efficiency.

@Ulysses Honestly, the moment it was leaked that it would have 12 SMs (1536 CUDA cores, 48 tensor cores, 12 RT cores), that was the moment most people realized that using Samsung 8N was simply not feasible. It would be too big, run too hot, and consume power like nothing. The various other discoveries like using an 8-core Cortex A78C and having 12GB of LPDDR5X capable of 120GB/s bandwidth just solidify that Samsung 8N is dead for this project, and that doesn't include various LinkedIn profiles that tie in work on the T239 SoC with work on Nvidia Grace, which is also made on TSMC 4N.

The idea of simply downclocking to handle thermals and power consumption like what they did with the Tegra X1 doesn't apply here because they aren't using an off-the-shelf SoC that they contracted out late in its development. They are making a custom one, which is up to Nintendo specifications from start to finish. Plus being Ampere-based, the lowest clock possible is 420 Mhz by design, and that's being under the mark of peak efficiency where they can't lower the voltage any further,. A lot of things are against Samsung 8N from a logical point of view. We can't consider them being late to use TSMC 4N either because that was first made available in 2020 when the first known tapeout of the T239 was in 2022 prior to the Linux commits.

But to answer your question about node change, it takes years because they have to follow procedure from start to finish, even with a similar design.

@TSR3 There was an article back in '22 by SemiAnalysis that while TSMC 4N was more expensive per wafer than Samsung 8N by 2.2x, it was also 2.7x more dense (more chips per wafer), effectively concluding it was cheaper per chip with TSMC 4N than Samsung 8N. Then there's the situation of yields, where TSMC 4N had far better yields than Samsung 8N, so more valid chips could be made with TSMC 4N. Granted, Samsung could sell by working chip than by wafer to get an overall lower price, but we're talking about making a chip for Nintendo's next console that is being prepped for pushing 100+ million units. I think even with a cheaper price per chip with Samsung, they couldn't afford not having enough supply. Take PS5, for instance. Chip production initially saw only a 50% yield, so a LOT of unusable chips were made.

With TSMC increasing prices of the 4N process node lately, that won't affect Nintendo's decision to use it because they've already locked it in at least since 2022 with the first tapeout and there is continued references to the T239 since then. If they had changed the chip, then the "T239" would be defunct and a new name would take its place. That's how Nvidia operates. And maybe they were able to lock in a price for the chip with TSMC long before any of this price increases happened. Who knows. All I know is that Samsung 8N is simply not feasible with what we know.

@Discostew 👍🏻 👍🏻 (that’s a like for both posts above. NL don’t have yeahs on their forums, you know what I mean 🤓)

@Discostew Thanks for sharing your perspective in-depth! Now I feel a lot more confident that Switch 2 will go with TSMC's 4N node. Take my thumbs up as well 🤓👍

@TSR3 @Ulysses
👍

I haven't read through this entire thread to know what has or hasn't been discussed, but I have been on the Famiboards for almost 2 years where they speculated and discovered info on the Switch 2, so if some of this information I'm about to present has already been discussed beforehand, then forgive me.

• SoC: Nvidia T239
• CPU: 8-core Cortex A78C (around 3x stronger than the A57 in Switch on a per-core, same clock basis)
• GPU: Ampere-based with 12 SMs (1536 CUDA cores, 48 tensor cores, 12 RT cores), includes some backported Lovelace features
• RAM: 12GB LPDDR5X @ 7500 MT/s (120GB/s)
• Display: 8-inch, 1080p
• Storage: 256GB eUFS 3.1 (peak 2.1GB/s)
• Extras: Hardware FDE block (File Decompression Engine), AV1 encode/decode support, Clock gating

Certainly some other things I'm missing, but those are much of the big ones. The clocks for the CPU and GPU have not been leaked, though one might assume the test suites from the Nvidia leak may give an idea of the GPU, where is gave the following while assigned with DLSS Performance Mode.

• 660 Mhz (4.2W) - 1080p
• 1.125 Ghz (9.3W) - 4k
• 1.38 Ghz (12W) - 4k

Given what we know of the number of CUDA/shader cores the GPU has, we can determine the raw performance of these profiles, being 2.0 TFlops, 3.45 TFlops, and 4.2 TFlops respectfully. Do note these were with DLSS Performance Mode, so maybe the power parameters include active use of tensor cores? We really don't know. There is something to mention here. DLSS is not as active as people might think. On a 4090, for instance, the tensor cores when used for DLSS to hit 4k pushes on average less than 1% utilization. Scaling that to the levels of the T239 would be in the realm of 20%.

So assuming that, what is happening with the other 80%? General-purpose. Based on the Ampere whitepapers, it would seem that whenever an FP16 operation is used, it's not handled by the GPU shaders, but by the tensor cores. If I'm correct, that means the numbers given here are technically not including FP16, and that using FP16 operations alongside FP32 will boost the raw performance altogether. Even with just a 25% utilization for general purpose, that would theoretically push the numbers to 2.5 TFlops, 4.31 TFlops, and 5.25 TFlops respectfully. But, let's assume from here on out that tensor cores aren't used for general purpose.

So back to DLSS. This imo is really going to make Switch 2 shine, especially in portable mode. Recall the mention of the profile for 1080p, which involved DLSS Performance Mode. This means the native render is at a 50% scale on both axis, so the native render would be 540p. Now, that sounds pretty low, right? Would the image quality suffer greatly? You may look at something like Series S using FSR, and how pretty terrible the upscaled quality looks, so one might think the same for DLSS. But, let me show you what DLSS 3.5 is capable of with an example. Doom Eternal, upscaled from 360p to 1080p using DLSS Ultra Performance Mode.

(When viewing the video, set the quality to max at 4K for the highest bitrate and the least number of compression artifacts. The video was recorded at 1080p, converted to 4K, and uploaded because uploading of a lower resolution results in a much lower bitrate, making it look bad)

Now, imagine that on an 8-inch screen if you're not using one already to view it. Heck, I watched it on my 65-inch TV, and it still looked pretty decent. From someone more knowledgeable than myself, they put the comparison of DLSS to FSR in this manner. For what FSR can do, DLSS can do better one major level lower (like Quality to Performance, 67% to 50%), and the lower the target resolution, the further ahead DLSS goes over FSR with upscaling it. So, take something like 1440p and upscaling to 4k using FSR Quality Mode (66%). DLSS would be able to at least match the resulting quality using Performance Mode (50%) with a 1080p resolution. 1080p is ~56% of 1440p. So you can imagine how much less a load on resources would be with DLSS to produce at least an equivalent result. To make matters worse for FSR, what DLSS does with dedicated hardware, FSR has to use existing GPU resources originally exclusive to rendering, ultimately reducing the render budget and overall power available.

So, in a case for Switch 2 portable mode, where using DLSS Performance to target 1080p with a clock of 660Mhz that pushes 2.0 TFlops, that's 2.0 TFlops shoved into a 540p native render. That may be much more than enough for portable mode, but Switch will also have an advantage elsewhere. Ray tracing, which will also run on dedicated hardware in Switch 2. I'll let your imagination run wild on that.

Putting less work on "raster" performance and more into "compute" performance is what Nvidia has been doing since 2019 with the Turing architecture. I honestly feel Switch 2 is going to be a beast. The only area where it may not push hard enough is in CPU power. Sure, the CPU that Switch 2 will have will run laps around what PS4 and other 8th-gen home consoles had, but it's unlikely to match 9th-gen home consoles in that regard. But, compared to Switch and 8th-gen home consoles, the gap is going to be much smaller thanks to using 8 cores rather than 4.

For handheld mode, I'd love to see a video on 720p vs 1080p, is 1080p really needed in handheld given the size of the screen? For many games would 720p @ 60FPS be better than 1080p @ 30FPS? If there had to be a trade off?

@MarioBrickLayer
It's also, fairly reliably, rumoured to be an 8" screen so it's also a fair bit bigger. What we're going to end up with is a screen that has over double the number of pixels than the Switch but with a pixel density that's about in line with the Switch Lite

Does it need 1080p at 8"? Honestly, probably not. But I think with the amount of power it has behind it should be able to comfortably drive that resolution. Especially with the assistance of DLSS. And with how fast screens have improved over the last few years why not?

@skywake if it can be done, it should, but without being technical (so this may be incorrect) I assumed there would be a trade off with battery life and/or frame rate. If 720p means a rock solid 60FPS for a 3D game like Zelda or COD and gives us an extra 45 minutes of battery then I would choose that every time over 1080p. It may be frame rate is a bigger draw on the battery than resolution, I'm not sure.

@MarioBrickLayer
Would depend on the game but I will say that in theory DLSS will change the curve somewhat. A long while ago I posted some images of graphs showing estimates of how much DLSS would "cost" vs natively rendering. It was a model of it so obviously was super simplified and real software has a lot more variables.... but.... it was good as an approximation

I still have that spreadsheet, if I plug in your theoretical 30fps at 720p then I get:
480p native at over 120fps (again, model, unlikely because of other bottlenecks)
480p upscaled to 720p at 100fps
480p upscaled to 1080p at 80fps
720p native at 60fps
720p upscaled to 1080p via DLSS at 46fps
720p upscaled to 1440p vis DLSS at 39fps
1080p native at 27fps
1440p native at 15fps (RIP)

so in your theoretical example where they can hit 60fps at 720p? I could imagine them doing a dynamic resolution scaling of something between 480p and 720p and pushing to 60fps at 1080p. Or if it's a 90hz screen possibly running at 45fps and doing 720p upscaled to 1080p. There are options

Imagine if NL had Likes for forum posts... Oh boy! I would have so many ❤ 0

@skywake you should probably put a spoiler tag on there.

This is crazy stuff though, do we know anything about more specs from this?

@PikminMarioKirby The specs leaked at same time, I’d put insert them myself but I don’t know how to do spoiler tags 😂

@FragRed [ spoiler ] [ /spoiler ] without the spaces, the spoiler text goes in between the two spoiler things like this

@PikminMarioKirby
Not much that isn't already known I don't think. USB port on the top, a clearer idea of how the magnets work, confirmation that the old JoyCon won't be physically compatible. Also the space they've left for the game card is suspiciously similar in size to the current Switch so I think the kind kills a lot more of the doubt that might've existed about that

...... if there was any doubt that we're close to reveal before

@skywake So that seems to be hinting at Backwards Compatibility? That was my most wanted feature, very glad that seems to be there. I wonder if controllers from Switch will still be compatible even if they can't connect.

If we weren't getting a reveal soon, we probably are now

@PikminMarioKirby

@FragRed That is crazy, and I don't really understand a lot of it, but perhaps that is a good thing so I'll be more surprised lol
does the 256 GB mean we're getting that much storage? 2100 MB/s seems pretty fast, how does that compare to Switch?