StarTopic Future Nintendo Hardware & Technology Speculation & Discussion |ST| (Read the staff posts before commenting!)

[ILikeFeet]

Not exactly doubting your original post, I agree with most of it but I still have some questions in my mind which I'll get to later.
First off, a bit of historical context to explain why I agree Nintendo picking 8nm as a node makes no sense:

Back in 2015, Switch's Tegra X1 chip release date, it's manufacturing process was 20nm. Taking a look at other mobile SoC's released at a similar timeframe, 20nm seems to be a common node, for example: the Snapdragon 810. But some were already being manufactured under 14nm like the Exynos 7420.
Of course, in 2015 if we look at any budget phone, we'll see that ~28nm was widely used.
But the Switch had a 2015 chip for a product released in early 2017, a year in which 10nm was becoming more common with chips like the Exynos 8895 on the higher end and 14nm on mid range phones with chips like the Exynos 7885 and Snapdragon 821.
Sure, you would still see 20+nm phones releasing in 2016 through 2017 in the low to very low end of the market but overall, 14nm and under was quite common. Now, one could question Nintendo on why they picked 20nm in 2017 instead of shipping 16nm early. And well, I think this relates to how the hardware was taped out for mass production back in a time where 20nm was the best option available to them.

Now, the more recent context for mobile chips released from 2022 through mid 2023 is: "low end" to mid range phones commonly use 5nm SoCs eg: Exynos 1330 and 1280. But, higher end actually expensive phones are shipping with 4nm chips such as: the Snapdragon 8 gen 2 and Exynos 2200.
But those are reaally fresh phone releases. I'm talking < 4 months ago. If we take a look just a while back, 7nm was quite common in some mid-range Sammy and Qualcomm chips like the Snapdragon 865.

In conclusion:
Considering previous historical context but, taking Thraktor's efficiency analysis into account, I'm not really questioning whether or not sub 8nm is what Nintendo will pick. Instead, I'm wondering whether or not they'll go 7nm with a die shrink later to 5nm or even 4nm or, if they'll wait until manufacturing under 5nm becomes cheaper.
But I'm starting to believe that they'll go with 5nm anyways because I believe they're aware of what happened to some 2017 switches by now.
And considering the hardware tapeout for drake took place around 2022 then I think there's no doubt.

But I'm still skeptical about how high they'll clock the SoC.

the TX1 as a barometer is gonna give you bad data because the situation is different. Nintendo didn't pick 20nm, Nvidia did. remember that the TX1 was not a gaming product, it was a multi-purpose chip that was meant to go into everything, from cars, to android set top boxes, to game consoles. they settled on 20nm because it was the next step after 20nm, but they and the rest of the industry realized 20nm was trash. but Nvidia bought a fuckload of capacity that they would have paid penalty for if Nintendo didn't step in.

Drake is not the same as it's made, ground up, for Nintendo at Nintendo's behest. you can't die shrink a 7nm down to 4nm like you could 20nm to 16nm/12nm. 16nm/12nm are greatly improved 20nm while 5nm is not the same as 7nm. it might cost Nintendo and Nvidia more money to design for 7nm and then design for 4N

 

[Tentacle-tropes]

Okay, thank you.

Now. Why? What’s the motivation to show off dev kits during an event where it could be leaked or maybe even stolen? Why not do it in a completely confidential event where no consumers would even know about it?

This event is confidential since most consumers don’t even realize the backroom deals that go on. I doubt it would have a chance to be stolen unless they physically bring it (even then), the amount of people will almost assuredly be small, & plenty of NDAs will abound. The motivation of doing such is to prep key 3rd parties about the next device so they have software within the launch window

 

[AshiodyneFX]

This event is confidential since most consumers don’t even realize the backroom deals that go on. I doubt it would have a chance to be stolen unless they physically bring it (even then), the amount of people will almost assuredly be small, & plenty of NDAs will abound. The motivation of doing such is to prep key 3rd parties about the next device so they have software within the launch window

Wouldn't security slightly slacking off would still be somewhat of a risk, why not have it be held privately entirely even if a small chance exists? also, cant they prep 3rd parties by sending them the SDK individually?

 

[Dakhil]

I imagine Nintendo's most trusted third party developers are already informed about devkits well in advance. So I imagine if Nintendo does talk about devkits during Gamescom 2023, Nintendo's probably talking to other third party developers.

 

[Tentacle-tropes]

Wouldn't security slightly slacking off would still be somewhat of a risk, why not have it be held privately entirely even if a small chance exists? also, cant they prep 3rd parties by sending them the SDK individually?

Why would security slack off for a location no one knows but the people attending the meeting? These backroom deals/talks are essentially private events. They can but as noted above this would be most likely be for other 3rd parties & retailers. This is what E3 was typically for & why you got in ages past what is essentially a quarterly breakdown from said company.

 

[AshiodyneFX]

Why would security slack off for a location no one knows but the people attending the meeting? These backroom deals/talks are essentially private events. They can but as noted above this would be most likely be for other 3rd parties & retailers. This is what E3 was typically for & why you got in ages past what is essentially a quarterly breakdown from said company.

Guess I misunderstood how far away it would be from the actual event, my bad, still, I don't know why they wouldn't just send the SDK to devs right away or do a private meeting away from Gamescom, if the security was that tight people wouldn't be expecting it to leak around then

 

[AshiodyneFX]

I imagine Nintendo's most trusted third party developers are already informed about devkits well in advance. So I imagine if Nintendo does talk about devkits during Gamescom 2023, Nintendo's probably talking to other third party developers.

who does nintendo usually prioritize for dev kits?

 

[Dakhil]

who does nintendo usually prioritize for dev kits?

Probably most major Japanese AAA third party developers (e.g. Capcom, Sega, Square Enix, etc.) and some major Western AAA third party developers (e.g. Mojang Studios, WB Games, etc.).

 

[ILikeFeet]

who does nintendo usually prioritize for dev kits?

that's confidential, but the biggest third parties are the obvious ones. Ubisoft has been the source of leaks in the past. when you have 20K+ employees, that's to be expected

 

[Ashee]

August this year; assuming that the reason Nintendo is going to Gamescom this year is to show off the new dev kits in private.

This would mean that we will have the new console by next year right? ( i guess half year or even holidays)

Im debating getting an oled switch like in 2 days or waiting for the new cosole (no, i dont want to wait an entire year)... i was sold back when the switch trailer was released but i didnt have the money, now i have but i dont want to buy a console and then in 6 months have a new gen (is not that easy to sell an used console here and if a new oled cost us$350, here it cost us$610, like 3-4 salaries)

 

[Look over there]

@Thraktor
I was looking at the official numbers. I actually didn't know that in game, the GPUs can go above them.

 

[kasparov77]

When the idea comes up, you’d have to ask yourself, “why would Nintendo actually do this? What’s the benefit? They are paying for more silicon, yet they aren’t even using it?

Didn't the Tegra X1 come with 4 cores software binned? was there any confirmation as to whether or not they were actually present on silicon?

 

[Dakhil]

Didn't the Tegra X1 come with 4 cores software binned? was there any confirmation as to whether or not they were actually present on silicon?

Nvidia did disable the 4 Cortex-A53 cores on the Tegra X1, but permanently.

A die shot of the Tegra X1 shows the 4 Cortex-A53 cores being present in silicon.

 

[Deleted member 887]

.

In conclusion:
Considering previous historical context but, taking Thraktor's efficiency analysis into account, I'm not really questioning whether or not sub 8nm is what Nintendo will pick. Instead, I'm wondering whether or not they'll go 7nm with a die shrink

Fortunately for us, we have versions of Ampere on both 8nm and 7nm processes, and while those versions of Ampere are different, it seems unlikely that 7nm is a viable option. But also TSMC is trying to eliminate the various 7nm class nodes they still have, leaving them for legacy customers.

But I'm still skeptical about how high they'll clock the SoC.

I mean, welcome to the loop

@Thraktor's analysis is pretty solid. I also tend to be pessimistic, but I don't think anything about his technical analysis is wrong. I think there are two sets of considerations.

The first is that, while Nintendo will get the best bang for buck at the peak efficiency, there is still some bang to be found below that. Orin has some profiles in the 400Mhz range, and Orin Nano even has a 300Mhz profile - though it should be noted the 300Mhz profile is documented to be saving no power over the 400 Mhz profile. Nintendo may choose to get some marginal savings, if they think they can get away with it.

The second is that Thraktor's power curves are probably pessimistic. Every reason we have to imagine that Drake's curves will be different than Orin's favors Drake. Drake is on more advanced node. Orin has power-sucking double rate tensor cores. Drake has additional power saving tech that Orin doesn't have.

@Thraktor's argument is that a 12 SM design seems to indicate a design that achieves its performance goal by going wide, and a wide design can really take advantage of peak efficiency, which is correct. My argument is that a wide design also sees the best gains from below peak efficiency, because marginal savings from each SM add up.

Ultimately, the clock settings will be decided by the hardware team and the software team duking it out, not by a designer working on the chip in abstract. Hardware will give EPD every drop of truly free performance they can - and free meaning not just free battery, but free in terms of the thermal design and the noise profile - and EPD will ask for all of the performance they can get. And every time hardware senses that EPD isn't using perf they've been given, hardware will cut back. And every time EPD senses there is over engineering they will take advantage.

Everything we do here is just trying to predict the result of that back and forth based on the properties of the technology being chosen. I personally tend to low ball my estimates for purely political reasons. There are less technical folks here who set their expectations based on us prattling along, so I tend to look at things from a "worst case" perspective, plus "what would maximizing battery life look like" perspective.

What's wild is that my prediction and @Thraktor's look so different on paper, but I think we basically agree on everything

 

[kasparov77]

Nvidia did disable the 4 Cortex-A53 cores on the Tegra X1, but permanently.

A die shot of the Tegra X1 shows the 4 Cortex-A53 cores being present in silicon.

Not saying they would repeat the mistake of shipping silicon with part of the hardware unusable again but...

 

[iag]

Any chance [redacted] comes with support to racing wheels, like Logitech’s, Thrustmaster’s, etc? Not many racing games in Switch but with the increase in power, maybe [redacted] will see the release of some of them.

 

[Sin Drive]

because the jump in graphical/techinal power between PS3/PS4 was considerable

What the hell are you on about?

 

[Deleted member 887]

Didn't the Tegra X1 come with 4 cores software binned? was there any confirmation as to whether or not they were actually present on silicon?

Nvidia did disable the 4 Cortex-A53 cores on the Tegra X1, but permanently.

A die shot of the Tegra X1 shows the 4 Cortex-A53 cores being present in silicon.

And TX1+, the revision of TX1 built specifically for Nintendo, removed them. Remember, while Nintendo was involved in the last stage of TX1 design, the chip was designed for Google, not Nintendo.

 

[TLZ]

I love the Switch OS, other console's OSs are needlessly busy with the worst kind of feature creep.

Switch OS is pretty snappy, responsive, and simple. I was surprised at how quickly I could capture 30s clips and share them too.

I do miss the catchy music though... And it's a shame how terribly the shop runs.

It's super bare bones and bland. I'd much rather Wii U's stylistic look. Even Wii's.

 

[TLZ]

Is it salt or sodium chloride?

Huh? Surely you understand what I mean. One can simply mean the same game made on last gen hardware first, maybe even no changes made for current gen just through BC, and the other made on current gen with significant changes made for the last gen port to be able to run natively. Maybe even made from scratch by another team because of the different architecture and hardware. They're not similar.

 

[kasparov77]

It's super bare bones and bland. I'd much rather Wii U's stylistic look. Even Wii's.

WiiU's aesthetics minus the slow ass transitions between apps.

 

[ItWasMeantToBe19]

The degree of cross-gen on PS4/PS5 often just was a boost from 30 FPS to 40 FPS and a resolution boost from 1080p to 4K.

Many games went much further with their PS5 versions, but even God of War Ragnarok did very little other than being 4K/40 and that was the most important game released on PS5.

 

[Tentacle-tropes]

Guess I misunderstood how far away it would be from the actual event, my bad, still, I don't know why they wouldn't just send the SDK to devs right away or do a private meeting away from Gamescom, if the security was that tight people wouldn't be expecting it to leak around then

People are expecting leaks because they hope Nintendo is gonna reveal it there. Typically the leaks would come from press material &/or show floor material. Nintendo will send their closest partners the SDK but for those not as close, haven’t worked with Nintendo, or want to pitch a project; a trade show event is the best place to catch them since if they are going you can typically try to schedule something.

 

[TLZ]

I have no idea what I'm doing wrong

You're not doing anything wrong. It's the game.

 

[JoshuaJSlone]

And it’s not that I think demand will be less than 20 million, it’s that I think supply is going to be less than 20 million. 20m is a difficult number to hit year one, it’s actually very rare to hit 20 million year one.

We of course don't get updates for every day of the year, but so far all the fastest sellers seem to have managed right at 17m in one calendar year.

 

[ILikeFeet]

Not saying they would repeat the mistake of shipping silicon with part of the hardware unusable again but...

Again, the TX1 is an off the shelf product. You can't use it as a precedence because the situation is different

 

[Pokemaniac]

When do you guys expect leaks of the actual hardware to happen? like the console T239 is going in I mean, say PS5 dev kit style leak, or like a switch lite factory shell leak, etc etc, also is it possible to find stuff about T239 software through a data mine of a pre-existing or upcoming game? say parameters for BC or something along those lines, or is that completely impossible? I am new, sorry if either was answered already and I missed it

Leaks are random occurrences. You can't really predict when they're going to happen.

It is plausible that information could be datamined from a game, but most of the obvious forms of that are things that Nintendo can fairly trivially check for and deny during the certification process, so it's fairly unlikely to actually happen before the hardware is announced.

i said this all along but seems like the Zelda delay was to release or at least have it be an option as releasing as a Switch Pro/2/Boogaloo launch title. Hardware ends up being delayed but Zelda is already delayed at this point so the extra time is used for 'polish'. money is on the enhanced version being deep in development at that point, maybe they moved onto DLC too as suggested.

Or the game just needed polishing. It's a big and complicated game, with a lot of stuff that could go wrong.

 

[ReddDreadtheLead]

Didn't the Tegra X1 come with 4 cores software binned? was there any confirmation as to whether or not they were actually present on silicon?

Not software binned, they were physically there. They just were turned off/fused off.

There were issues and no one used the A53s. They had no purpose for a gaming console to begin with, as the big.LITTLE setup of the A57+A53 meant one or the other, not both at once. And that would be disastrous for a console, where it’s a constant fixed environment, and would limit some functions like suspension.

The Tegra X1 though wasn’t necessarily designed with a console such as the Nintendo switch in mind; it was designed for something like the Google Pixel, a android tablet, a set top TV streaming box, etc.

The switch is, for all intents and purposes, using an off the shelf component due to the circumstances that Nintendo were in at the time. Those A53s weren’t disabled at Nintendo’s request though, they were disabled from all existing Tegra X1 SOCs.

And the Nintendo switch is of course a dedicated gaming console with its own closer to the metal game development with a thin layer. Lest not forget that important element.

The degree of cross-gen on PS4/PS5 often just was a boost from 30 FPS to 40 FPS and a resolution boost from 1080p to 4K.

Many games went much further with their PS5 versions, but even God of War Ragnarok did very little other than being 4K/40 and that was the most important game released on PS5.

Not to be rude, but I’m unsure of the significance of this comment or who you were replying to. Are you saying to expect a cross gen patch to be higher res and higher FPS? Or that we shouldn’t get out hopes up for a lot more than that?

I’m perfectly fine with a patch for a higher resolution and a higher framerate.

 

[Hermii]

the TX1 as a barometer is gonna give you bad data because the situation is different. Nintendo didn't pick 20nm, Nvidia did. remember that the TX1 was not a gaming product, it was a multi-purpose chip that was meant to go into everything, from cars, to android set top boxes, to game consoles. they settled on 20nm because it was the next step after 20nm, but they and the rest of the industry realized 20nm was trash. but Nvidia bought a fuckload of capacity that they would have paid penalty for if Nintendo didn't step in.

Drake is not the same as it's made, ground up, for Nintendo at Nintendo's behest. you can't die shrink a 7nm down to 4nm like you could 20nm to 16nm/12nm. 16nm/12nm are greatly improved 20nm while 5nm is not the same as 7nm. it might cost Nintendo and Nvidia more money to design for 7nm and then design for 4N

People at the time said it would be too expensive to shrink from a planar process to a finfet process, so Mariko had to be a new chip. Yet that's exactly what they did. And back then, Nvidia didn't use AI for downshrinks.

I'm no expert, but I feel the die shrink issue is overblown. Maybe they won't change foundries, but within the same foundry I feel downshrinks are doable.

 

[ILikeFeet]

People at the time said it would be too expensive to shrink from a planar process to a finfet process, so Mariko had to be a new chip. Yet that's exactly what they did. And back then, Nvidia didn't use AI for downshrinks.

I'm no expert, but I feel the die shrink issue is overblown. Maybe they won't change foundries, but within the same foundry I feel downshrinks are doable.

Maybe it is, maybe it isn't. Regardless, I don't expect it. Path of least resistance would be to just make a bespoke product for the smallest node as cheaply as possible and 4N is looking to be that

 

[Triaura]

Given that Capcom advised Nintendo to increase memory of the Switch I think they are involved in [redacted] development or at least Nintendo asked about their opinion.

 

[Hartmann]

Macronix plans to release 96-layer 3D Nand Flash for the game console manufacturer this year.

旺宏3D Nand 96層產品今年可放量 主供遊戲機業-MoneyDJ理財網

旺宏(2337)今年雖面臨大環境消費景氣差，持續減產，但在研發進度以及研發費用上，公司則未縮手，公司目前新產品的開發如進度進行，3D Nand flash 96層產品已陸續給客戶sample，主攻遊戲機產業，等對方驗證後，...

www.moneydj.com

 

[ReddDreadtheLead]

Macronix plans to release 96-layer 3D Nand Flash for the game console manufacturer this year.

旺宏3D Nand 96層產品今年可放量 主供遊戲機業-MoneyDJ理財網

旺宏(2337)今年雖面臨大環境消費景氣差，持續減產，但在研發進度以及研發費用上，公司則未縮手，公司目前新產品的開發如進度進行，3D Nand flash 96層產品已陸續給客戶sample，主攻遊戲機產業，等對方驗證後，...

www.moneydj.com

Nintendo… unless I missed something

So, a 128GB cart? Jesus Christ. Must cost a fortune if ever actually used.


Edit: Wait, no. Does this mean the 32GB cards are at the price similar to the 16GB carts?

 

[Bonejack]

Nintendo…

So, a 128GB cart? Jesus Christ. Must cost a fortune if ever actually used.

And yet. it still couldn't store Jedi Survivor fully on it!

(This is a dig on the games wasteful size hunger, not about the size of the cart!)

;]

 

[ReddDreadtheLead]

And yet. it still couldn't store Jedi Survivor fully on it!

(This is a dig on the games wasteful size hunger, not about the size of the cart!)

;]

It’s not about the physical size of the cart, it’s about how much storage you have and how you use it.


Or something like that

 

[boodream]

Doesn't the fact that the TX1 was not designed specifically for Nintendo while its successor for the next Switch will be necessarily imply a significant price increase compared to the advantageous conditions they were able to obtain at the time? And if so, to what extent will this cost increase impact the price of the Switch's successor?

 

[pelusilla6]

And yet. it still couldn't store Jedi Survivor fully on it!

(This is a dig on the games wasteful size hunger, not about the size of the cart!)

;]

It's criticisable and can probably be reduced with optimization but you can't compare 4k textures and hours of cinematics againts 720p textures with five short cinematics.

 

[Stefan51278]

People at the time said it would be too expensive to shrink from a planar process to a finfet process, so Mariko had to be a new chip. Yet that's exactly what they did. And back then, Nvidia didn't use AI for downshrinks.

I'm no expert, but I feel the die shrink issue is overblown. Maybe they won't change foundries, but within the same foundry I feel downshrinks are doable.

The 20nm-node was used for the TX1 and the first years of the Switch because it did cost Nvidia basically nothing to do this. They paid for a lot of 20nm-capacity and had to buy wafers or pay a penalty-fee, but the node didn't perform well, especially not for bigger chips like desktop-class-GPUs. The former 28nm-node got more refined with every year used (like the famous 14nm-node of Intel which was used for over half a decade because their 10nm-node was overambitious) and was replaced after many years by the new finfet-16nm-node. Nvidia already declared these wafers as a loss in their financial numbers before the deal with Nintendo was made, which made the revenue with Big-N a big success for the balance of the company. I think they made Nintendo an offer they couldn't refuse.

The shrink for the Switch was happening after only two years because they wanted to harden the design against homebrew anyway and Nvidia ran out of 20nm-wafers faster than anyone could have predicted in the planning-phase of the hybrid console.

Doesn't the fact that the TX1 was not designed specifically for Nintendo while its successor for the next Switch will be necessarily imply a significant price increase compared to the advantageous conditions they were able to obtain at the time? And if so, to what extent will this cost increase impact the price of the Switch's successor?

Nintendo was very cautious after having a flop with a custom-designed WiiU which cost them a lot of penalty-fees for not buying as much chips as planned. They didn't want to bet their bank deposit on another custom-design in that situation and that would also have made the dreadful WiiU-lifecycle even longer. After the success of the Switch they got their Mojo back.

 

[Angel Whispers]

Given that Capcom advised Nintendo to increase memory of the Switch I think they are involved in [redacted] development or at least Nintendo asked about their opinion.

Partner feedback is a general thing in product development. Gearbox told Sony they “don’t listen”, and were “done” if they went for 4GB RAM instead of 8 in the PS4, while Epic advised Microsoft to have 512MB RAM in the X360 instead of 256. So, each platform has a case, and making the call to go with more resulted in an 80m+ sales for X360, 116m sales for PS4, and, by the time it ends production, 140m+ for the Switch. I’ve been increasingly confident of 16GB RAM in the successor - While 12 would still be welcome, it’s hard to ignore what Nvidia has literally put out there regarding this. More so, for something expected in 2024 or later, but also because more flagship-specced phones will have that, AND it will be imperative for RT, DLSS, and the SoC’s feature sets.

 

[Cuzizkool]

While we’re talking about TX1 and TX1+, does anyone know what GFLOPS TX1+ is capable of if Nintendo hadn’t throttled it for battery life? As in, if they kept the same power profile as V1 Switch. Come to think of it I don’t know the GFLOPS of V1 Switch either. The difference wouldn’t be huge I imagine, but I’m just curious. LPDDR4X would have made a difference too I bet.

 

[Ghostsonplanets]

People at the time said it would be too expensive to shrink from a planar process to a finfet process, so Mariko had to be a new chip. Yet that's exactly what they did. And back then, Nvidia didn't use AI for downshrinks.

I'm no expert, but I feel the die shrink issue is overblown. Maybe they won't change foundries, but within the same foundry I feel downshrinks are doable.

16nm was a direct evolution from 20nm, even though it changed transistor type. We're also on a whole different situation from back then. Wafer, design, verification, etc, have skyrocket compared to 16nm. Hence why some of us think Nintendo would rather go all out first, so that they can get great battery life and small enough chip that it can be reused for the Lite later, without needing a shrink that would be costly.

Doesn't the fact that the TX1 was not designed specifically for Nintendo while its successor for the next Switch will be necessarily imply a significant price increase compared to the advantageous conditions they were able to obtain at the time? And if so, to what extent will this cost increase impact the price of the Switch's successor?

Yes. But at the same, this is a different situation from Wii U. Back then, Nintendo was in a dire situation and didn't knew if the concept would be successful, hence why they needed costs to be as low as possible. Now they have proven concept and can spend a lot more as they know it will be successful. A custom SoC will also support their ambitions from 1P to 3P.

While we’re talking about TX1 and TX1+, does anyone know what GFLOPS TX1+ is capable of if Nintendo hadn’t throttled it for battery life? As in, if they kept the same power profile as V1 Switch. Come to think of it I don’t know the GFLOPS of V1 Switch either. The difference wouldn’t be huge I imagine, but I’m just curious. LPDDR4X would have made a difference too I bet.

TX1 was capable of 1TFLOP FP16 and 512 GFLOPs FP32.

 

[ReddDreadtheLead]

While we’re talking about TX1 and TX1+, does anyone know what GFLOPS TX1+ is capable of if Nintendo hadn’t throttled it for battery life? As in, if they kept the same power profile as V1 Switch. Come to think of it I don’t know the GFLOPS of V1 Switch either. The difference wouldn’t be huge I imagine, but I’m just curious. LPDDR4X would have made a difference too I bet.

OG switch had frequencies of 307.2, 384 and 460.8 MHz for portable mode, with 768MHz for docked mode.

So that would translate to 157.28GFLOPs, 196GFLOPs and 235.93GFLOPs.

Docked would be 393.21GFLOPs

I believe you can run the docked GPU speed in portable on the V2/OLED units and get a better battery life still, but once you include increasing the CPU speed you get a similar battery life to the OG switch.



When nvidia had the Tegra X1 posted originally, it was able to deliver 512GFLOPs of FP32 (and 1024GFLOPS of FP16) running the GPU at 1GHz and a CPU that can hit up to 2.2GHz..


But that throttled and Nintendo went with a safer option.



TX1+ though can deliver beyond that as a peak of 649GFLOPs FP32 (1298 GFLOPs FP16) at 1267MHz and a CPU that can reach 2.1GHz.

 

[Thraktor]

I think as long as TSMC's N6 process node or newer is used for the fabrication of Drake, I think I'm happy.

I should say that I'd be happy with pretty much any manufacturing process, as we already know how big the GPU is, and it's unlikely the clocks would be below around 500MHz on any plausible process for reasons discussed, so my performance expectations wouldn't change much either way. Whether they achieve that by using TSMC 4N, as I expect, or they manage to make it work on N6 or Samsung's 5nm/4nm processes through some kind of witchcraft doesn't make much difference to me.

Nintendo doesn't prioritize high performance as much as their competitors do, but that doesn't mean they're aiming for low performance. The TX1 is one of the fastest chips you could find at that power envelope in 2016, and is already capable of outperforming a PS3 in most real world workloads that get thrown at it. It stands to reason they'll probably maximize performance for their power budget again, as being within porting range, even if just barely, from PlayStation and Xbox has served them well and helped to really sell the Switch as a console you can take with you rather than just a portable.

I think "performance for their power budget" is the key. This isn't a case of Nintendo prioritising high performance, it's a case of Nintendo prioritising power efficiency. From what I can tell, it looks like Nintendo have been making every decision possible to squeeze as much power efficiency out of the chip as they can.

As I mentioned recently, it seems that when Nintendo were presented with the possibility of a GPU somewhere between Ampere and Ada, they took exactly two features from Ada: the TSMC 4N manufacturing process and some kind of additional clock gating called FLCG. I don't think it's a co-incidence that those are, as far as I can tell, the only parts of Ada that improve power efficiency over Ampere. That they skipped over the improved RT cores, SER, etc., but went for a clock gating approach that Nvidia doesn't even bother documenting tells us a lot about where their priorities were.

The size of the GPU also tells us they're going for a wide-and-slow design to improve power efficiency, potentially the widest and slowest (and therefore most efficient) they can manage. I know a lot of people are looking at devices like the Steam Deck and Ally to try to discern what is or isn't possible for the new Switch, but these devices really aren't designed for efficiency, they're designed for performance and cost with efficiency a distant third. This is particularly evident in their narrow-and-fast GPU designs.

Consider the Steam Deck GPU, which achieves 1.6Tflops by clocking 512 "cores" at 1.6GHz on TSMC N7. For comparison, the Xbox Series X GPU, which uses a similar RDNA2 architecture also on TSMC N7, is clocked at 1.825GHz. The Steam Deck is clocking its GPU at up to 88% the clock of a big, power hungry home console, despite being a portable, battery powered device. It actually clocks higher than the Series S's 1.565GHz.

The Ally is even worse. With a 768 "core" RDNA3 GPU on TSMC N4, claims a peak clock of 2.7GHz. That's a higher clock speed than AMD use on their 300W+ desktop RDNA3 GPUs. It's the very definition of narrow-and-fast design, and it shouldn't be surprising that the device chugs battery if you want clocks anywhere near 2.7GHz.

Of course you can't compare FLOPS directly across architectures, but the distinction is pretty clear when you consider that [redacted]'s GPU has 1536 "cores". That's three times the size of the Steam Deck GPU and twice the size of the Ally GPU, for a device that's almost certainly targeting much lower power consumption. While Valve and Asus went for small GPUs clocked almost as high as you'd expect in a stationary home console, Nintendo have gone with a GPU almost as big as one you'd expect in a home console, and will be clocking it much lower. The former is what you get if you design for performance and cost, the latter is what you get if you design for efficiency.

 

[Ghostsonplanets]

TX1+ though can deliver beyond that as a peak of 649GFLOPs FP32 (1298 GFLOPs FP16) at 1267MHz and a CPU that can reach 2.1GHz.

Imagine Switch Pro in 2021 with 7nm TX1 2.5GHz CPU and 1.2 GHz GPU, 8GB RAM and with DLSS Chip as SCD communicating through USB 2

Spoiler

just joking hahaha

What could have been.....

 

[ReddDreadtheLead]

I should say that I'd be happy with pretty much any manufacturing process, as we already know how big the GPU is, and it's unlikely the clocks would be below around 500MHz on any plausible process for reasons discussed, so my performance expectations wouldn't change much either way. Whether they achieve that by using TSMC 4N, as I expect, or they manage to make it work on N6 or Samsung's 5nm/4nm processes through some kind of witchcraft doesn't make much difference to me.



I think "performance for their power budget" is the key. This isn't a case of Nintendo prioritising high performance, it's a case of Nintendo prioritising power efficiency. From what I can tell, it looks like Nintendo have been making every decision possible to squeeze as much power efficiency out of the chip as they can.

As I mentioned recently, it seems that when Nintendo were presented with the possibility of a GPU somewhere between Ampere and Ada, they took exactly two features from Ada: the TSMC 4N manufacturing process and some kind of additional clock gating called FLCG. I don't think it's a co-incidence that those are, as far as I can tell, the only parts of Ada that improve power efficiency over Ampere. That they skipped over the improved RT cores, SER, etc., but went for a clock gating approach that Nvidia doesn't even bother documenting tells us a lot about where their priorities were.

The size of the GPU also tells us they're going for a wide-and-slow design to improve power efficiency, potentially the widest and slowest (and therefore most efficient) they can manage. I know a lot of people are looking at devices like the Steam Deck and Ally to try to discern what is or isn't possible for the new Switch, but these devices really aren't designed for efficiency, they're designed for performance and cost with efficiency a distant third. This is particularly evident in their narrow-and-fast GPU designs.

Consider the Steam Deck GPU, which achieves 1.6Tflops by clocking 512 "cores" at 1.6GHz on TSMC N7. For comparison, the Xbox Series X GPU, which uses a similar RDNA2 architecture also on TSMC N7, is clocked at 1.825GHz. The Steam Deck is clocking its GPU at up to 88% the clock of a big, power hungry home console, despite being a portable, battery powered device. It actually clocks higher than the Series S's 1.565GHz.

The Ally is even worse. With a 768 "core" RDNA3 GPU on TSMC N4, claims a peak clock of 2.7GHz. That's a higher clock speed than AMD use on their 300W+ desktop RDNA3 GPUs. It's the very definition of narrow-and-fast design, and it shouldn't be surprising that the device chugs battery if you want clocks anywhere near 2.7GHz.

Of course you can't compare FLOPS directly across architectures, but the distinction is pretty clear when you consider that [redacted]'s GPU has 1536 "cores". That's three times the size of the Steam Deck GPU and twice the size of the Ally GPU, for a device that's almost certainly targeting much lower power consumption. While Valve and Asus went for small GPUs clocked almost as high as you'd expect in a stationary home console, Nintendo have gone with a GPU almost as big as one you'd expect in a home console, and will be clocking it much lower. The former is what you get if you design for performance and cost, the latter is what you get if you design for efficiency.

Having read this, I’m honestly curious what the max permissive clock speeds would have been for a SoC like T239 in say, some set too box.

ie, what is the highest clock frequency possible before it ruins the bed.

Like, uh, 2GHz GPU and 3GHz for the CPU perhaps?

Imagine Switch Pro in 2021 with 7nm TX1 2.5GHz CPU and 1.2 GHz GPU, 8GB RAM and with DLSS Chip as SCD communicating through USB 2

Spoiler

just joking hahaha

What could have been.....

Utter chaos

 

[Foltzie]

Having read this, I’m honestly curious what the max permissive clock speeds would have been for a SoC like T239 in say, some set too box.

ie, what is the highest clock frequency possible before it ruins the bed.

Like, uh, 2GHz GPU and 3GHz for the CPU perhaps?


Utter chaos

We’ll find out when/if the T239 based Shield launches.

 

[ILikeFeet]

Having read this, I’m honestly curious what the max permissive clock speeds would have been for a SoC like T239 in say, some set too box.

ie, what is the highest clock frequency possible before it ruins the bed.

Like, uh, 2GHz GPU and 3GHz for the CPU perhaps?

I'd say as high as Orin, but if Drake is on 4N, then the limits are higher. but we're seeing that Lovelace is locked down to not perform any better across top end wattages without shunts. probably the same here

 

[Hermii]

Imagine Switch Pro in 2021 with 7nm TX1 2.5GHz CPU and 1.2 GHz GPU, 8GB RAM and with DLSS Chip as SCD communicating through USB 2

Spoiler

just joking hahaha

What could have been.....

I was with you until you brought up SCD lol.

 

[Hermii]

We’ll find out when/if the T239 based Shield launches.

Or if Nvidia fucks up their security.

 

[MP!]

I have no idea what I'm doing wrong

Yeah the game's framerate drops frequently and in all sorts of places... I actually don't think it's related to rendering exactly... I think it's streaming issues. The framerate drops as it tries to stream in assets and load them to memory... it hardly affects gameplay or becomes unplayable... though obviously responsiveness becomes an issue the lower the framerate. I think I've seen some sort of dynamic resolution scaling or something occur which is why I think it's a streaming issue... cause any gpu related frame dips would be covered up by dynamic resolution adjusting to maintain framerate. But I'm no expert here. I think "chugging" as you've described it is a bit more than what this is...(I'd describe chugging as the game literally stopping and trying to catch up multiple times... when in all circumstances I've witnessed the game is still fully playable just real FRAMEY) Oh ... I play mainly docked on my OLED model.
I can't wait for new hardware haha

anyway that's my guess.