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

[Kise Ryota]

That's another nail in the coffin for 8nm isn't it?

Even at like 300mhz, 8nm woudnt be that?

I think the lowest clock we could use on that Orin power tool was 420.75MHz. If we use this clock and something like 1.1GHz for the CPU, maybe we could be looking to a total system power draw of 13w (with 7.7w being from cpu+gpu at high loads).

With a 40Wh battery it could achieve 3h in a more demanding game (and close to 6h in a less demanding indie)

I really don't think it's impossible SEC 8nm when we are open to the idea of a bigger console. The possibility of a relatively bigger screen caught me off guard tbh (initially I was expecting OLED's size)

Anyway, it still doesn't make much sense to me to go with 8nm, but I don't rule out the possibility either (just as I don't rule out the possibility of it being in any other node other than SEC 8nm or TSMC 4N)

 

[chocolate_supra]

They don't. The gamepad is 23% heavier.

wait, is that with joycon detached or something?

I just picked up my Switch in one hand and my gamepad in the other and they feel extremely close in weight. Matter of fact when I began bobbing them up and down I'd say the Switch is the one that feels a bit heavier, by just a tiny amount. I know that's not a scientific form of measurement but I assume I'd feel nearly a quarter difference in weight.

 

[Kise Ryota]

wait, is that with joycon detached or something?

Switch v1/v2 with joy-cons = 398g (oled = 420g)
Gamepad = 491g

 

[Aether]

Part of the reason the gamepad doesn't feel much heavier is cause it's less dense.
Your mind expects something of its size to be heavier then it is, also the weight distribution that shifts less extreme since it's over a bigger volume.

 

[chocolate_supra]

Switch v1/v2 with joy-cons = 398g (oled = 420g)
Gamepad = 491g

weird..

Part of the reason the gamepad doesn't feel much heavier is cause it's less dense.
Your mind expects something of its size to be heavier then it is, also the weight distribution that shifts less extreme since it's over a bigger volume.

Yeah that's why I had them both on my palms and sorta bobbed them up and down in unison to see which one felt like it was weighing my hand down more, and that was the Switch.

Oh well

 

[Guaraná]

the thread got unpinned?

 

[ReyVGM]

the thread got unpinned?

Read mod post #93, 221.

 

[theguy]

The mods just didn’t want us to reach page 2000 too quickly, so this was a very good move by the team

 

[Hermii]

The mods just didn’t want us to reach page 2000 too quickly, so this was a very good move by the team

My bet for 3000 pages by launch is in shambles.

 

[ReddDreadtheLead]

I think it was just the original development of T234 back in 2020-2022. Orin NX and Nano are not new T234 dies, they're just binned chips, so no work on the memory subsystems of those SoCs was being done in 2023 or "currently," any more than it was for T239. So that part of the profile should not read be literally. The only Tegra "currently" in development is Thor, and it's not 8nm.

Personally I think the node discussion should have died when we started getting reports of the performance being impressive, probably even exceeding the expectations people were setting for ray tracing. We already know Nintendo won't release it with like one hour of battery life, and we can be reasonably sure they won't make an enormous handheld. So what's there even to worry about?

Ok but will it be ugly?

We need a sexy switch 2, let’s worry about the aesthetic instead.

 

[ReddDreadtheLead]

I can't speak for cost, but N3E won't allow for a big perf increase.
I may be sort of repeating myself? Similar response as I made here.

Oh, one more point that I haven't mention recently. I'd like to re-iterate my opinion that given 128-bit LPDDR5 or LPDDR5X, I think that N5 family is just about right. I think that as far as what N3E allows for, you'd probably need the standardization of LPDDR5T to truly realize the potential. And as far as I'm aware, LPDDR5T is still SK Hynix-specific.

Speaking of:

 

[ILikeFeet]

epic is working on a "neural post processing" for UE5, whatever that is. wonder if that can use tensor cores

 

[Bonejack]

epic is working on a "neural post processing" for UE5, whatever that is. wonder if that can use tensor cores

Wouldn't it be easier for them to partner with nVidia? (If DLSS is already supported in the engine ignore this, i legit don't know.)

 

[ns156]

Wouldn't it be easier for them to partner with nVidia? (If DLSS is already supported in the engine ignore this, i legit don't know.)

DLSS 3.5 is in UE5 AFAIK.

 

[Angel Whispers]

A good tagine Nintendo could do with Switch 2 is call it "the most powerful smallest console". They obviously cannot compete spec wise but this is something Nintrndo could do if they wanted to have fun with it.

Except the trouble with that is it’s always short-lived. I preferred Iwata saying they went with industry-leading chips, and Nintendo not making a fuss about it in the marketing - That was true, and still holds water now because of the innovation in SoC design, and the games which came to the platform. Going the power route leads to an inevitability that your competitors will catch up and surpass you, and it’s not even the most important thing. For the Switch, it was getting XB1/PS4 performance and games at a tenth of the power consumption, and up to 1TF via mixed precision. For the next platform, it will be getting PS5/XS performance and games at a fraction of the power consumption - But it would be smarter engineering and further innovation in this space that gets there. Which is why I’m always saying “smarter engineering, NOT raw power, will ultimately prevail”. It isn’t appreciated enough just how far ahead of the curve they are on this. Nobody ever bought consoles on the back of their power, anyway, as they’ve never eclipsed PCs. It’s always been about the games, not the specs. It never helped the XSX or the XB1X. It was an insignificant factor in the PS4’s success (being $100 cheaper than the XB1 at launch was more important), and it wasn’t a factor for the PS3 (X360 had the better performance than it on the most popular non-exclusive titles).

So far, they’ve made a statement to developers by showing the Matrix Awakens demo on their hardware, barely any loading times, and building on their relations - BG3 is a lock-in, unless you believe Nintendo employees make crack-filled donuts for Larian to get a place in the credits. CP2077 is a genuine possibility given that Nvidia has worked closely on it. So, showing the games you can expect to play on it matters more. It’s also telling that reports say “comparable to PS5/XSX” - It doesn’t matter that this can be through DLSS or better optimisation. Consumers don’t care if it’s cosmetic or native, and they won’t call it “cheating” like some folk in old and new old places, or other such wildness.

 

[Skittzo]

Wouldn't it be easier for them to partner with nVidia? (If DLSS is already supported in the engine ignore this, i legit don't know.)

DLSS is fully supported, neural post processing would seem to be a different thing or their own version of DLSS-type features.

 

[Ghostsonplanets]

epic is working on a "neural post processing" for UE5, whatever that is. wonder if that can use tensor cores

Neural Post-Processing? Whatever it is, I guess it's another step into the Neural Rendering dream?

 

[FireGrey]

Something that bugs me about T239 is the fact it’s using Ampere rather than Lovelace.
Afaik it wouldn’t be numbered 239 anymore if it was on Lovelace.
Does anyone with more knowledge know if it would be easier to “port” ampere to Samsung 4/5nm than it would be to TSMC 4nm?
In my mind the only way things make sense is if it was easier for them to update Ampere to use a smaller Samsung node than design a Lovelace chip for Nintendo.
Apparently Samsung 4nm has equal if not better yield than TSMC 4nm at the moment, and some analysts are saying it will have more production capacity and might entice Nvidia and others to use it over TSMC’s.
Unlike Samsung 8nm vs TSMC 7nm there doesn’t seem to be as great a transistor density gap between the companies for 5nm and 4nm, so I’m sure Samsung’s smaller nodes will meet performance/efficiency expectations.

 

[ILikeFeet]

Wouldn't it be easier for them to partner with nVidia? (If DLSS is already supported in the engine ignore this, i legit don't know.)

not the same thing. this is a function being developed in UE5

Neural Post-Processing? Whatever it is, I guess it's another step into the Neural Rendering dream?

don't know. it was updated 2 days ago

 

[redmutineer75]

Something that bugs me about T239 is the fact it’s using Ampere rather than Lovelace.
Afaik it wouldn’t be numbered 239 anymore if it was on Lovelace.
Does anyone with more knowledge know if it would be easier to “port” ampere to Samsung 4/5nm than it would be to TSMC 4nm?

Well, all of Nvidia's Lovelace products are on TSMC 4N, a node custom-made for Nvidia, so I imagine that changing to Samsung would be quite logistically and bureaucratically challenging for what would likely be only marginal improvements.

 

[ILikeFeet]

Something that bugs me about T239 is the fact it’s using Ampere rather than Lovelace.
Afaik it wouldn’t be numbered 239 anymore if it was on Lovelace.

Ampere and Lovelace aren't that much different. Lovelace's additions are largely cache, RT core additions, and OFA improvements

 

[mjayer]

Ampere and Lovelace aren't that much different. Lovelace's additions are largely cache, RT core additions, and OFA improvements

I think maybe his point is that if it is indeed on TSMC 4N, wouldn't that essentially make it a Lovelace product and it would have a different number than 239?

Hmm.

 

[kvetcha]

I think maybe his point is that if it is indeed on TSMC 4N, wouldn't that essentially make it a Lovelace product and it would have a different number than 239?

Hmm.

I would assume not unless it also integrated all the other changes included in Lovelace.

 

[redmutineer75]

I think maybe his point is that if it is indeed on TSMC 4N, wouldn't that essentially make it a Lovelace product and it would have a different number than 239?

Hmm.

Well, there is a difference between "essentially Lovelace" and actual Lovelace. For example, they might not include the enhanced OFA if it's not going to use frame gen. Nvidia wouldn't consider it to be a proper Lovelace chip if it doesn't have the whole enchilada.

 

[Dekuman]

Well, there is a difference between "essentially Lovelace" and actual Lovelace. For example, they might not include the enhanced OFA if it's not going to use frame gen. Nvidia wouldn't consider it to be a proper Lovelace chip if it doesn't have the whole enchilada.

Situation might be similar to X1, which was a maxwell class chip but had Pascal features because it came out relatively late in the Maxwell cycle.

In a way Maxwell /Pascal is a lot like Ampere/Lovelace with the latter leaning more on node shrink and refinements over the prior generation.

 

[Hermii]

Well, there is a difference between "essentially Lovelace" and actual Lovelace. For example, they might not include the enhanced OFA if it's not going to use frame gen. Nvidia wouldn't consider it to be a proper Lovelace chip if it doesn't have the whole enchilada.

There are also probably hundreds/ thousands of smaller differences, that aren't apparent before you really dig deep.

 

[Cybergatuno]

Just throwing a bit of tinder into the fire ...

Samsung Says It Will Beat TSMC's 4nm Production Capacity

Samsung's progress in the 4nm chip-making process puts it on par with TSMC. Learn how Samsung aims to surpass TSMC's production capacity

www.gizchina.com

Samsung has been working on improving its 4nm chip-making process for a while now, and it seems that the company is making significant progress. According to Sammobile and 9to5google, Samsung’s 4nm process is now on par with TSMC’s, and the company is confident that it will beat TSMC’s 4nm production capacity soon. Samsung’s steady progress in improving its 4nm chip yield may earn the company new orders from key clients such as Qualcomm and NVIDIA.

It has always been TSMC 4nm VS Samsung 8nm.
We may now have a new possibility, even if remote.

 

[Deleted member 887]

Even at like 300mhz, 8nm woudnt be that?

The two wild cards are "we don't know battery size" and "Drake has Lovelace's power saving tech." I personally don't believe that's enough, but I will admit that it's not 100% certain.

 

[Dakhil]

Just throwing a bit of tinder into the fire ...

Samsung Says It Will Beat TSMC's 4nm Production Capacity

Samsung's progress in the 4nm chip-making process puts it on par with TSMC. Learn how Samsung aims to surpass TSMC's production capacity

www.gizchina.com

It has always been TSMC 4nm VS Samsung 8nm.
We may now have a new possibility, even if remote.

The only process node newer than Samsung's 8N process node that Nvidia's rumoured to be using from Samsung is Samsung's 3 nm* process node.

The reason why I mention this is because outside of Nvidia only using TSMC's 20 nm* process node for the Tegra X1, Nvidia generally secures process node capacity for more than one product.

* → a marketing nomenclature used by all foundry companies

 

[ILikeFeet]

that reminds me, DF's video on downclocking a 2050 still isn't up yet. wonder if Richard is doing more extensive testing

 

[Alovon11]

Something that bugs me about T239 is the fact it’s using Ampere rather than Lovelace.
Afaik it wouldn’t be numbered 239 anymore if it was on Lovelace.
Does anyone with more knowledge know if it would be easier to “port” ampere to Samsung 4/5nm than it would be to TSMC 4nm?

Well for some historical context, Lovelace is pretty much AmpereV2, heck there are data bits that highly indicate that it was internally called AmpereV2 at some point before being named to Lovelace.

So the difference between Ampere and Lovelace as per NVIDIA is seemingly all the feautres over Ampere together.

So that is

• Shader Execution Reordering on the CUDA Cores
• Opacity Micromap Engine and Displaced Micro-Mesh Engines introduced into the RT Cores, along with a 2x IPC (Triangle Intersection Speed for the case of the RT cores) increase
• 2x IPC increase on the Tensor Cores and whatever a FP8 Transformer Engine is on Hopper?
• Higher Speed and Dense Optical Flow (Full Resolution) capable Optical Flow Accelerator
• Way Bigger L2 Cache
• Power Optimization Features at the Architecture Level (First-Level Clock Gating)
• TSMC 4N for overall power savings and clock increases (Unsure how much of the Tensor and RT Core's 2x raw speed increase int the ADA Whitepaper may be due to this or not though as Lovelace in general has 2x the clockspeeds of Ampere and the RT Performance doesn't seem indicative of a 2x IPC Increase as per usual???)

So, for reference to what we know about T239, it would only cover some of these via technicality? Even assuming TSMC 4N

• SER on the Shaders?
  • No, they're Ampere's CUDA cores from what we can tell.
• OME and DMME in the RT cores?
  • No, they're Ampere's RT cores from what we can tell.
• FP8 Transformer Engine in the Tensor Cores?
  • No, they're Ampere's Tensor cores from what we can tell.
• Dense Optical Flow OFA?
  • Yes actually, the T239 has the Orin OFA, which is capable of Dense Optical Flow, however the caveat is we don't know if the Orin OFA is the same as the Lovealce OFA, and even if it isn't we don't know the speed of the OFA, for all we know considering the application of Orin in Automotive tasks, the OFA may be faster at Optical Flow than Lovelace for all we know. And equally it could be slower.
• Bigger L2?
  • Nope, the L2 is consistent with Ampere's Scaling for the most part, and having theoretical CPU L3 access and maybe a System-Level Cache probably wouldn't tip the balance too much?
    • Although an interesting wrinkle is the Latency stall issue on Ampere may be alleviated a tad by using LPDDR which is what Cache is used primarily to alleviate as more data could be stored in a bigger cache, preventing the GPU from stalling by making a very long (relative to Cache or DDR latency) wait to call the GDDR for something.
• Power Optimization Features at the Architecture Level?
  • Actually First-Level Clock Gating is in T239, as for other features that Lovelace added, that's unknown there.
• TSMC 4N?
  • Well..that's what we're debating right?
  • IMHO 4N makes sense from the production end of things as it likely will be cheaper than Samsung 8N due to wafter cost/density equations making cost in scale lean towards TSMC 4N being cheaper (Denser Nodes = Less Wafers to get the same Volume of Product is the summarization of it).
    • Like, considering NVIDIA is already getting a good deal on Samsung 8N and that deal price is the one that comparison is being made off of. It'd take NVIDIA doing the Foundry Equivalent of Highway Robbery to make Samsung 8N cheap enough to outweigh all the other benefits TSMC 4N would bring.

 

[ShadowFox08]

Except the trouble with that is it’s always short-lived. I preferred Iwata saying they went with industry-leading chips, and Nintendo not making a fuss about it in the marketing - That was true, and still holds water now because of the innovation in SoC design, and the games which came to the platform. Going the power route leads to an inevitability that your competitors will catch up and surpass you, and it’s not even the most important thing. For the Switch, it was getting XB1/PS4 performance and games at a tenth of the power consumption, and up to 1TF via mixed precision. For the next platform, it will be getting PS5/XS performance and games at a fraction of the power consumption - But it would be smarter engineering and further innovation in this space that gets there. Which is why I’m always saying “smarter engineering, NOT raw power, will ultimately prevail”. It isn’t appreciated enough just how far ahead of the curve they are on this. Nobody ever bought consoles on the back of their power, anyway, as they’ve never eclipsed PCs. It’s always been about the games, not the specs. It never helped the XSX or the XB1X. It was an insignificant factor in the PS4’s success (being $100 cheaper than the XB1 at launch was more important), and it wasn’t a factor for the PS3 (X360 had the better performance than it on the most popular non-exclusive titles).

So far, they’ve made a statement to developers by showing the Matrix Awakens demo on their hardware, barely any loading times, and building on their relations - BG3 is a lock-in, unless you believe Nintendo employees make crack-filled donuts for Larian to get a place in the credits. CP2077 is a genuine possibility given that Nvidia has worked closely on it. So, showing the games you can expect to play on it matters more. It’s also telling that reports say “comparable to PS5/XSX” - It doesn’t matter that this can be through DLSS or better optimisation. Consumers don’t care if it’s cosmetic or native, and they won’t call it “cheating” like some folk in old and new old places, or other such wildness.

Huh. The Switch was getting ports of xbone and PS4 games, but not at their performance. No amount of hardware engineering put it on equal footing as those two. There's a significant gap between them.

And it definitely isn't getting 1 tflops mixed precision. You must have gotten that 1 tflops from fp16 numbers, but No game could be played 100% in fp16, mot to mention, the GPU is only clocked at 768Mhz.

 

[Angel Whispers]

Something that bugs me about T239 is the fact it’s using Ampere rather than Lovelace.
Afaik it wouldn’t be numbered 239 anymore if it was on Lovelace.
Does anyone with more knowledge know if it would be easier to “port” ampere to Samsung 4/5nm than it would be to TSMC 4nm?
In my mind the only way things make sense is if it was easier for them to update Ampere to use a smaller Samsung node than design a Lovelace chip for Nintendo.
Apparently Samsung 4nm has equal if not better yield than TSMC 4nm at the moment, and some analysts are saying it will have more production capacity and might entice Nvidia and others to use it over TSMC’s.
Unlike Samsung 8nm vs TSMC 7nm there doesn’t seem to be as great a transistor density gap between the companies for 5nm and 4nm, so I’m sure Samsung’s smaller nodes will meet performance/efficiency expectations.

Lovelace is probably more likely at this point, tbqh. We still don’t know a lot about what it is. The leaked info, while official, is still a foundation, the basement of expectations for performance. We don’t know if it’s a placeholder or a dev kit for developers to work with - The fact that the demo showed last month was on “targeted specs” is the giveaway that not everything is nailed on, and reports on DLSS 3.5 would allude to this SoC being developed with other Lovelace products. The SoC does have some elements, after all!! BTW, there were whispers about the next Nintendo platform having a Lovelace GPU in it as far back as 2021, but many conflated that with the idea of a Switch “Pro”. But the key is not to fall into the trap of worrying. Ampere is still more modern than the AMD-powered systems, while Lovelace builds on that. The SoC can have elements of both because ultimately, it’s a CUSTOM chip.

 

[JoshuaJSlone]

400g is too heavy for a handheld in my experience. I also think the Switch pro controller is slightly heavier than ideal.

Is it at all possible that the 8 inch screen could be smartphone aspect ratio (about 20.5:9)? That would make it about the same height as the standard Switch's screen, but wider. The extra screen real estate on either side could be used for virtual buttons unique to each game.

I don't think it's likely at all that they'd either give portable/docked different aspect ratios for games, or give them different sets of controls--except where motion/touch necessitates it.

Foundry claims don't combine perf improvement and power improvement together.
It's perf improvement for the same power. Or, it's power improvement for the same frequency. Not both at the same time.

It's not an all of one or all of the other choice. If Switch had started with Mariko rather than Erista, they could've chose to make it 10% faster and still use less power.

 

[Angel Whispers]

Huh. The Switch was getting ports of xbone and PS4 games, but not at their performance. No amount of hardware engineering put it on equal footing as those two. There's a significant gap between them.

And it definitely isn't getting 1 tflops mixed precision. You must have gotten that 1 tflops from fp16 numbers, but No game could be played 100% in fp16, mot to mention, the GPU is only clocked at 768Mhz.

Don’t read too much into it - Some games will play the same, others won’t. It could perform and run any of the games those systems had in some capacity, not always with 1 for 1 parity, and in the rare case, better, even. Perhaps I could’ve articulated that better, but I wasn’t all there, so, please, allow it. The SoC was revealed and pitched as I described it - I quoted that at face value. The 1TF was FP16, hence the “via mixed precision, and I was still accurate because I said “Up To”. Also, the gaps weren’t significant to most people, and software sales proved that time and time again. We’re talking Dreamcast/PS2 to GameCube/XBox comparisons at worst. That’s as much as I’ll write on this.

 

[Mr.Gamerson]

The Switch is already much larger than previous Nintendo handheld game consoles, and that size change, coupled with the ubiquity of smartphones, has made the Switch much less common on Japanese public transportation than the DS and 3DS. Despite this, Switch has sold 30 million units in Japan, and the Switch OLED is currently the best-selling Switch model in Japan.
So there's no reason why the Switch 2 can't be even bigger.

If that 7.91 lcd rumor is true, then Nintendo likely agrees with you, as do i. I assume the next system will likely be larger than the Oled, but not much so, and that's completely fine, considering it's a hybrid system. it will also allow the next Lite console to be more attractive to users due to the size difference between the two being more pronounced.

 

[Concernt]

Lovelace is probably more likely at this point, tbqh. We still don’t know a lot about what it is. The leaked info, while official, is still a foundation, the basement of expectations for performance. We don’t know if it’s a placeholder or a dev kit for developers to work with - The fact that the demo showed last month was on “targeted specs” is the giveaway that not everything is nailed on, and reports on DLSS 3.5 would allude to this SoC being developed with other Lovelace products. The SoC does have some elements, after all!! BTW, there were whispers about the next Nintendo platform having a Lovelace GPU in it as far back as 2021, but many conflated that with the idea of a Switch “Pro”. But the key is not to fall into the trap of worrying. Ampere is still more modern than the AMD-powered systems, while Lovelace builds on that. The SoC can have elements of both because ultimately, it’s a CUSTOM chip.

T239 is in all likelihood the SOC of the next Nintendo console, and it's an Ampere+ design. Look at it this way, they already have testbeds in their SDK and their beefier Orin chips, so a bespoke testbed SOC would be an extreme expense that doesn't make sense business wise. The production timescale we see only makes sense if they intend to release it and release it as part of a Nintendo system. Lovelace did not redefine graphical computing, it's a refinement on Ampere. Relevant to gaming would be its improved OFA and thus frame generation capabilities, but it's recommended that such is only used with framerates targeting 80 or above (40 internally), something Nintendo likely has no interest in due to the low adoption of 120hz TV sets, and the fact their hardware is low power and may not benefit from it at all at the performance it targets.

Meanwhile the main benefit of Lovelace for a Nintendo console would be the smaller node and improved power efficiency; something that T239 on 4N would have.

Then we have to consider that, in general, it's far easier to develop software for something that's a SUBSET of your hardware capabilities rather than a SUPERSET. The way they have it set up, Lovelace GPUs are a superset of T239, and so make development easier than if they were not a superset. Meanwhile, T239 still inherits all RELEVANT benefits of Lovelace.

This "Ampere+" design is actually very smart, and on release next year it'll still be cutting edge in the realm of edge and mobile computing. Nobody else does or can have a 10W DLSS capable device.

 

[Lancelot]

This "Ampere+" design is actually very smart, and on release next year it'll still be cutting edge in the realm of edge and mobile computing. Nobody else does or can have a 10W DLSS capable device.

Admittedly... It doesn't really matter what's cutting edge or not at this point, this is something people don't seem to grasp at all. For all we know they could use LPDDR5x on this thing and it'll still be the best performing handheld ever for quite some time. Mobile GPUs are not desktop GPUs and linearly increasing their performance doesn't necessarily yield gains, memory bandwidth will and always has been the biggest limitation when dealing with handheld performance. Until we see an LPDDR5T/LPDDR6 device, the T239 will remain unmatched in its league no matter how much more powerful this hypothetical killer may be.

 

[Ghostsonplanets]

it's so insulting about Samsung foundry.
Remember, that 7LPH joke is about a further refinement of their 10nm generation node.

It isn't insulting and neither a joke. Samsung 8nm is the cheapest non-EUV node available and still desirable for many applications. Hence why Samsung investing into further and for specific clients refinements isn’t something out of extraordinaire. Nvidia 8N, going by timelines, is a custom DTCO onto Samsung 8LPP, which was 1st Gen 8nm. Samsung refined the node even further with 8LPU and another refinement that was made available after Nvidia 8N products which was 8LPA. Hence why I take power estimates of 8nm with caution, as we base it on Ampere 8N products but don't know if 8N is the most efficient custom 8nm. What if 8LPA, which according to Samsung consumes 15% less power than 8LPU, is a more efficient 8nm flavor than 8N? T239 could be fabbed using it.

Samsung Foundry is starved for clients and the idea of further 8nm refinement to meet Nintendo and Nvidia demands for T239 isn't something we can throw out of the window.

It's 50/50. It's either TSMC 4N or a Samsung 8nm flavour.

 

[Aostia82]

totally dumb question here

with DLSS being almost certain for Switch REDACTED...will this make developers focusing the rea calculation power on other aspects than the native resolution, trying to improve, for example, other performance aspects such as framerate?

 

[Lancelot]

totally dumb question here

with DLSS being almost certain for Switch REDACTED...will this make developers focusing the rea calculation power on other aspects than the native resolution, trying to improve, for example, other performance aspects such as framerate?

Well, with DLSS there's no need to run anything at native resolution so developers could indeed augment the capabilities of REDACTED as a result, more gpu power going towards materials, lightning and such. If they wanted to use the gains to push 60 FPS they could too, it's game dependent.

 

[Deleted member 887]

totally dumb question here

with DLSS being almost certain for Switch REDACTED...will this make developers focusing the rea calculation power on other aspects than the native resolution, trying to improve, for example, other performance aspects such as framerate?

Not dumb. That is, in fact, exactly what DLSS is intended to do

 

[Ghostsonplanets]

Something that bugs me about T239 is the fact it’s using Ampere rather than Lovelace.
Afaik it wouldn’t be numbered 239 anymore if it was on Lovelace.
Does anyone with more knowledge know if it would be easier to “port” ampere to Samsung 4/5nm than it would be to TSMC 4nm?
In my mind the only way things make sense is if it was easier for them to update Ampere to use a smaller Samsung node than design a Lovelace chip for Nintendo.
Apparently Samsung 4nm has equal if not better yield than TSMC 4nm at the moment, and some analysts are saying it will have more production capacity and might entice Nvidia and others to use it over TSMC’s.
Unlike Samsung 8nm vs TSMC 7nm there doesn’t seem to be as great a transistor density gap between the companies for 5nm and 4nm, so I’m sure Samsung’s smaller nodes will meet performance/efficiency expectations.

One of the bigger reasons why we're limiting ourselves to either 8N or 4N is that these are the nodes where Nvidia us making products. In case of Switch 2 not meeting Nvidia and Nintendo projected sales, Nvidia could repurpose the wafers into others products such as Gaming or DC GPUs, Tegra Orin, etc. If Nvidia were to choose a node like Samsung 5LPP/4LPP, in case Switch 2 sells lower than projected, Nvidia couldn't repurpose the wafers into others products and thus would need to pay a fine for underutilization.

Basically, for Nvidia, it makes more financial and logistical sense for them to fab T239 on a node they already use.

 

[Goodtwin]

It isn't insulting and neither a joke. Samsung 8nm is the cheapest non-EUV node available and still desirable for many applications. Hence why Samsung investing into further and for specific clients refinements isn’t something out of extraordinaire. Nvidia 8N, going by timelines, is a custom DTCO onto Samsung 8LPP, which was 1st Gen 8nm. Samsung refined the node even further with 8LPU and another refinement that was made available after Nvidia 8N products which was 8LPA. Hence why I take power estimates of 8nm with caution, as we base it on Ampere 8N products but don't know if 8N is the most efficient custom 8nm. What if 8LPA, which according to Samsung consumes 15% less power than 8LPU, is a more efficient 8nm flavor than 8N? T239 could be fabbed using it.

Samsung Foundry is starved for clients and the idea of further 8nm refinement to meet Nintendo and Nvidia demands for T239 isn't something we can throw out of the window.

It's 50/50. It's either TSMC 4N or a Samsung 8nm flavour.

@oldpuck has spoken on it before, but if its 8nm, there has to be power saving measures that we are not aware of and are not found in other 8nm processors. Its still hard to square everything because the processor would still be much larger at 8nm than 4N, resulting in less chips per wafer. There is a lot of extra things that need to be in place to make 8nm make sense for T239. Perhaps Samsung made Nvidia/Nintendo an extremely aggressive offer to get them to go with 8nm, giving them a low cost on the wafers and a low cost on the manufacturing cost. Maybe Nvidia came up with power savings tech that makes a significant difference. The problem here should start to be come obvious, in order for 8nm to work all these other things need to come to fruition. A ridiculous deal from Samsung and power savings tech that we haven't seen in other products all need to be there. On the flip of that, we have 4N sitting there as a possibility, and suddenly things just work with no extra work arounds. Power draw is where is needs to be with no additional mystery tech. Yields are great so cost per SOC is good and Nvidia already uses 4N for all their RTX40 graphics cards.

 

[ILikeFeet]

@oldpuck has spoken on it before, but if its 8nm, there has to be power saving measures that we are not aware of and are not found in other 8nm processors. Its still hard to square everything because the processor would still be much larger at 8nm than 4N, resulting in less chips per wafer. There is a lot of extra things that need to be in place to make 8nm make sense for T239. Perhaps Samsung made Nvidia/Nintendo an extremely aggressive offer to get them to go with 8nm, giving them a low cost on the wafers and a low cost on the manufacturing cost. Maybe Nvidia came up with power savings tech that makes a significant difference. The problem here should start to be come obvious, in order for 8nm to work all these other things need to come to fruition. A ridiculous deal from Samsung and power savings tech that we haven't seen in other products all need to be there. On the flip of that, we have 4N sitting there as a possibility, and suddenly things just work with no extra work arounds. Power draw is where is needs to be with no additional mystery tech. Yields are great so cost per SOC is good and Nvidia already uses 4N for all their RTX40 graphics cards.

8N is based on 8LPU if I remember correctly. 8LPA is a refinement on that but hasn't gone used. But 8LPA is from 2021, which is perfect timing for Drake development. We won't know completely until a product using it comes out.

 

[Ghostsonplanets]

@oldpuck has spoken on it before, but if its 8nm, there has to be power saving measures that we are not aware of and are not found in other 8nm processors. Its still hard to square everything because the processor would still be much larger at 8nm than 4N, resulting in less chips per wafer. There is a lot of extra things that need to be in place to make 8nm make sense for T239. Perhaps Samsung made Nvidia/Nintendo an extremely aggressive offer to get them to go with 8nm, giving them a low cost on the wafers and a low cost on the manufacturing cost. Maybe Nvidia came up with power savings tech that makes a significant difference. The problem here should start to be come obvious, in order for 8nm to work all these other things need to come to fruition. A ridiculous deal from Samsung and power savings tech that we haven't seen in other products all need to be there. On the flip of that, we have 4N sitting there as a possibility, and suddenly things just work with no extra work arounds. Power draw is where is needs to be with no additional mystery tech. Yields are great so cost per SOC is good and Nvidia already uses 4N for all their RTX40 graphics cards.

Agreed. I would add that one thing that people seem to forget is that Nvidia - Samsung agreement was rumored to be "pay for working dies only". Basically that Samsung Foundry would eat the losses on non-functioning dies. So the Samsung x NV deal is already aggressive as it is and throw off the economics of it a little.

 

[Goodtwin]

8N is based on 8LPU if I remember correctly. 8LPA is a refinement on that but hasn't gone used. But 8LPA is from 2021, which is perfect timing for Drake development. We won't know completely until a product using it comes out.

Oh don't get me wrong, I haven't completely wrote off the possibility. It just brings more questions that currently cant be answered. Would the wafers be the same for 8nm that Orin uses if they were to use 8LPA? There are a lot of problems that need solutions with 8nm, all the while 4N just works. Its not like 4N wasn't a thing in throughout Drakes development, Nvidia was very aware of this process node throughout its development. In a way, if its does end up being 8nm, its would be neat to see the creative ways that made it work.

 

[Kise Ryota]

How many tensor cores are needed to go from native 1080p to 2160p using DLSS' performance mode? Is there a way to calculate that?

Is the workload always the same? I mean, does it always need the same number of TC's to push 4 times more pixels (in the case of performance mode, for example) or it is something that will vary in a per game basis?

Also, is there a way to lower the quality of the DLSS to make it work with less TC's?

I imagine that the more TC's you have, the faster it will get the job done. But I'm curious to know how many TC's are needed to have DLSS working at its best quality in a frame time budget that is viable.

 

[Pokemaniac]

totally dumb question here

with DLSS being almost certain for Switch REDACTED...will this make developers focusing the rea calculation power on other aspects than the native resolution, trying to improve, for example, other performance aspects such as framerate?

In general, DLSS tries to reduce the GPU load by rendering fewer pixels and using a mixture of an ML model and a few other techniques to fill in the gaps. In a like for like comparison, this generally means an improved framerate, but games that aren't required to run without it are likely to bake any gains they get from that into their performance budget from the get-go.

 

[Ghostsonplanets]

Oh don't get me wrong, I haven't completely wrote off the possibility. It just brings more questions that currently cant be answered. Would the wafers be the same for 8nm that Orin uses if they were to use 8LPA? There are a lot of problems that need solutions with 8nm, all the while 4N just works. Its not like 4N wasn't a thing in throughout Drakes development, Nvidia was very aware of this process node throughout its development. In a way, if its does end up being 8nm, its would be neat to see the creative ways that made it work.

One thing that makes me very positive about thinking that Nintendo choose a very big, and agressive sized for a mobile part, GPU, was due to the fact they and Nvidia had insight into what 4N would enable them to do and at what power points.
1536 Shader Cores is very big for a tablet like SoC on 8nm. But on 4nm? It isn't at all and we even have comercially available mobile SoCs with 1536 Shader Cores, at reasonable 8 - 5W power envelopes, in the form of Snapdragon 8 Gen 2. The 8G2 is even clocked at the same clocks we predict T239 Portable Mode on 4N would clock, which is ~680MHz.