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

[Terrell]

Somehow everybody concluded that TX1+ is using 16nm.

It’s because it’s mostly confirmable, the SoC in Switches and Shield TVs from 2019 onward are a smaller die size with the exact same performance and the only node change possible from 20nm through TSMC without the SoC having a much smaller die for the same performance. Never mind that we can know this after the fact based on what node the SoCs are being fabricated on over at TSMC.

 

[ReddDreadtheLead]

Somehow everybody concluded that TX1+ is using 16nm.

I don’t think that’s a comparable scenario. With the Tegra X1+ it already has a point of comparison, the Tegra X1. It was physically smaller, cooler and at the same level of performance you could get a longer battery life with no major issues.

What would Drake have? It has no point of comparison, it can be assumed to be 7nm, or 6nm or 5nm or 4nm or a super custom 8N that is only for that SoC that nvidia didn’t use with the other SOCs that managed to give it +30-50% iso perf at the same power consumption, on the same node no less.


If the chip itself is absolutely tiny, like sub 100mm^2, it’s either 7 or 5. If it’s >160mm^2 then it would fall to the 8nm side.

100-130mm^2 it is probably 7nm.

 

[ReddDreadtheLead]

If the SSDs are already getting hot now, then it doesn't seem like there's much overhead left in terms of amount of watts that can be going into a drive of that form factor. So, if there's a goal of increasing raw sequential read in the following gen, one option is work off of presumed improvements in perf/watt of later generations, but not go anywhere near full throttle. Use some p-state in the middle. Alternatively, see how eUFS develops and go with that. Samsung did announce UFS 4.0 this year, claiming read speeds of up to 4.2 GB/s. Fast forward to the time of 'The Next Generation', maybe there's UFS 5.0 with a doubling to 8.4 GB/s. And UFS having the design principles it has, that should easily offer multiple GB/s per watt (I'm guessing that UFS 3.1 already should hit at least 2 GB/s per watt under typical operating conditions).

Speaking of heat in the big consoles, how's the RAM? GDDR7 should be doubling GDDR6's speeds while the change in signaling is alleged to reduce power draw by 25%, so it seems like double the bandwidth for 1.5x the power, before factoring in future node-related improvements. Can the console form factor support cooling full speed GDDR7?
And although it hasn't been stated yet, I think that it's likely that density is doubled. GDDR6 comes in 1 or 2 GB 32-bit chips, so we could be looking at 2 or 4 GB 32-bit chips for GDDR7. Now considering the potential slowdown in increase in raw sequential read, ~doubling the amount of RAM may not necessarily be enough for Next Generation experiences? There are options to RAM quantity (DDR and LPDDR should be beating GDDR handily in that area), but they do involve the tradeoff of relatively far less bandwidth.

(I've probably hinted at it a few times by now, but I am interested in a timeline where the set top boxes evolve towards something like, 256-bit LPDDR6/6X or something with fat amounts of RAM and get back under a hundred watts)

I think in this case for capacity, perhaps GDDR7W is where it would be headed? It seems like that reduces the physical footprint a bit but not of the SoC.


Idk it just seems like they have a lot of options for these consoles going forward if they want to. To make something around the same level of efficiency for example, a trade off can be supplanted here. You mentioned eUFS and that can work theoretically if they go with that high end memory as a trade off in power consumption. Maybe have multiple UFS embedded to supply the storage needs of the consoles.

Or maybe they’ll forgo this and allow these to get hot enough or draw more than their current which is what, like 200W?

 

[Z0m3le]

I don’t think that’s a comparable scenario. With the Tegra X1+ it already has a point of comparison, the Tegra X1. It was physically smaller, cooler and at the same level of performance you could get a longer battery life with no major issues.

What would Drake have? It has no point of comparison, it can be assumed to be 7nm, or 6nm or 5nm or 4nm or a super custom 8N that is only for that SoC that nvidia didn’t use with the other SOCs that managed to give it +30-50% iso perf at the same power consumption, on the same node no less.


If the chip itself is absolutely tiny, like sub 100mm^2, it’s either 7 or 5. If it’s >160mm^2 then it would fall to the 8nm side.

100-130mm^2 it is probably 7nm.

if the GPU is twice as efficient as Orin at the same configuration and clock, it really doesn't matter what node we think it is, the end result is that it performs like a 5nm Orin GPU. Shrug you want to assume it's 8nm, there is only so much that actually matters.

 

[ReddDreadtheLead]

if the GPU is twice as efficient as Orin at the same configuration and clock, it really doesn't matter what node we think it is, the end result is that it performs like a 5nm Orin GPU. Shrug you want to assume it's 8nm, there is only so much that actually matters.

So long as it has a moderately fast enough CPU idc lol

 

[Codename_Steam]

if the GPU is twice as efficient as Orin at the same configuration and clock, it really doesn't matter what node we think it is, the end result is that it performs like a 5nm Orin GPU. Shrug you want to assume it's 8nm, there is only so much that actually matters.

How big would Drake be on Samsung 8NM vs TSMC 4NM?

 

[ILikeFeet]

Guys, it's clearly Samsung 7nm

 

[mariodk18]

UHS-II and UHS-III are, as you point out, quite a bit pricier, as are other options other than UFS Card. M.2 storage has a power consumption and size problem, so I leave that out.

The main impediment for external card storage is, as @Pokemaniac suggested earlier, Nintendo putting in the legwork to establish the market for it that does not yet exist, no matter which microSD UHS-I alternative they choose. That is doable, especially for UFS for the reasons previously mentioned, and they absolutely should choose the option that is cheapest to the consumer, but it operates under an assumption.

That assumption being that, if internal storage and Game Card read speeds increase as we expect they may need to, Nintendo will want to maintain the ability for loading games off external storage and find a solution that grants them read speed parity. But they could just as likely revert to the idea of external storage being storage ONLY and require games to be played off internal storage only, bringing us back to the days of the Wii. Such a decision will ultimately come down to how much internal storage capacity they have, so as to prevent game juggling across storage when possible.

Nintendo has historically cheaped out on internal storage, but this device is likely to be sold out for quite a while and bought by enthusiasts at first, so I could see them going with one 256 GB SKU at launch for $449. Then maybe by holiday 23 or 24 they introduce a 128 GB SKU for $399. Maybe they do a deal with Samsung for UFS cards to get them at a good price.

 

[ReddDreadtheLead]

How big would Drake be on Samsung 8NM vs TSMC 4NM?

We already know some things about the SoC, so just to name a few:

L2$: 7.6mm^2

PCIe 16x 4.0 : 5.18mm^2

Digital Logic: 33.93mm^2

GPC: 51.9mm^2

This is already nearly 100mm^2 (98.61) using the rest of the ampere as comparisons. Even ORIN.

 

[Dakhil]

Guys, it's clearly Samsung 7nm

No, Samsung's 8LPP+ process node or Samsung's 10LPP++ process node or Samsung's 5LPE process node.

 

[ILikeFeet]

No, Samsung's 8LPP+ process node or Samsung's 10LPP++ process node or Samsung's 5LPE process node.

I thought there was still shrinkage going from 8nm to 5nm?

but hey, there's still samsung 4nm. it's good enough for Qualcomm low end, it's good enough for nintendo

 

[mariodk18]

So I was listening to various old podcasts that discussed Switch 2 rumors and I found another instance of Nate using the word "heavy" in the description of this video going over the Capcom data breach:

Capcom recently became the victims of a data breach & a mountain of information from the company -- ranging from personnel files to their software roadmap for 2021. With several unannounced titles mentioned in the files, as well as intents to bring announced games to other platforms, the leak is heavy; however, in the planning for 2021 releases, there is one noticeable omission: the Nintendo Switch Pro or 2021 revision. What does this omission mean? We discuss the leak and whether this is a red flag about any planning Nintendo hardware revision releasing next year.

Going by this usage of the word, it's clear to me that Nate views the word "heavy" in the same way as: significant, great, dense, profound, "relatively large extent and density"or "of great gravity or crucial import; requiring serious thought". I think Nate may have already cleared up his use of the word earlier, but I just wanted to help...reassure those who thought that "heavy" meant "bad news" or something of the sort.

 

[Terrell]

Nintendo has historically cheaped out on internal storage, but this device is likely to be sold out for quite a while and bought by enthusiasts at first, so I could see them going with one 256 GB SKU at launch for $449. Then maybe by holiday 23 or 24 they introduce a 128 GB SKU for $399. Maybe they do a deal with Samsung for UFS cards to get them at a good price.

In terms of the cards, I see them going to Western Digital/SanDisk, their current official partner for external storage, first, then maybe Samsung if SanDisk doesn't want to go for it.
Considering the severe reduction in NAND flash demand currently (projected to fall again by 20% in the final quarter of 2022), it puts Nintendo in a favourable position to negotiate supply contracts regardless of who supplies them with internal and external storage, as suppliers will be relieved to find a high-volume customer willing to soak up their NAND and lessen how much they have to wind down NAND production, so an internal eUFS 3.x at 128GB would probably be favourably priced at least for the first year or so, and by year 2 or later, economies of scale and reduced demand for eUFS 3 in favour of eUFS 4 will cushion any price increase seen for the internal storage in NAND demand sees a rebound.

It's a good time to be in the market for embedded or external flash storage, is what I'm saying.

 

[mariodk18]

In terms of the cards, I see them going to Western Digital/SanDisk, their current official partner for external storage, first, then maybe Samsung if SanDisk doesn't want to go for it.
Considering the severe reduction in NAND flash demand currently (projected to fall again by 20% in the final quarter of 2022), it puts Nintendo in a favourable position to negotiate supply contracts regardless of who supplies them with internal and external storage, as suppliers will be relieved to find a high-volume customer willing to soak up their NAND and lessen how much they have to wind down NAND production, so an internal eUFS 3.x at 128GB would probably be favourably priced at least for the first year or so, and by year 2 or later, economies of scale and reduced demand for eUFS 3 in favour of eUFS 4 will cushion any price increase seen for the internal storage in NAND demand sees a rebound.

It's a good time to be in the market for embedded or external flash storage, is what I'm saying.

I'm seeing a lot of, I assume, high-end smartphones with 256 GB of UFS 3.1 storage and 12 GB of RAM. And the One Plus 9? I believe I saw for about $500. Would getting a deal with Samsung for a good number of components be able to get that in Drake for $449 in May at launch?

 

[ezidro3]

A weird and quite frankly, unsubstantiated little theory:

Generally, each calendar year for the Switch usually ends with a media app launching:

• 2017: Hulu US (November 9th)
• 2018: YouTube (November 8th)
• 2019: N/A
• 2020: Funimation (December 15th)
• 2021: Pokémon TV (August 26th), Twitch (November 11th)
• 2022: Crunchyroll (February 17th), Hulu JP (April 27th).

4 out of 6 years isn't a terrible trend to go off of. I wouldn't be surprised if a media app launches alongside Drake touting 4K (and hopefully HDR) streaming. I think the most likely candidates are Peacock (owned by Universal, Mario movie is supposedly going on here after 45 days) and Disney+ (technically Disney already works with Nintendo through Hulu).

 

[Magic-Man]

A weird and quite frankly, unsubstantiated little theory:

Generally, each calendar year for the Switch usually ends with a media app launching:

• 2017: Hulu US (November 9th)
• 2018: YouTube (November 8th)
• 2019: N/A
• 2020: Funimation (December 15th)
• 2021: Pokémon TV (August 26th), Twitch (November 11th)
• 2022: Crunchyroll (February 17th), Hulu JP (April 27th).

4 out of 6 years isn't a terrible trend to go off of. I wouldn't be surprised if a media app launches alongside Drake touting 4K (and hopefully HDR) streaming. I think the most likely candidates are Peacock (owned by Universal, Mario movie is supposedly going on here after 45 days) and Disney+ (technically Disney already works with Nintendo through Hulu).

I'd be shocked if Peacock wasn't on Switch by the end of next year. Disney+ is a good candidate too. Netflix and HBO Max are lost causes though, unfortunately.

 

[ezidro3]

I'd be shocked if Peacock wasn't on Switch by the end of next year. Disney+ is a good candidate too. Netflix and HBO Max are lost causes though, unfortunately.

Looks like you can access HBO Max through Hulu so I guess that's a solution for that. Agreed on Netflix though, it would've come out way sooner

 

[Dakhil]

I thought there was still shrinkage going from 8nm to 5nm?

but hey, there's still samsung 4nm. it's good enough for Qualcomm low end, it's good enough for nintendo

There's probably still shrinkage probably thanks to the usage of EUV lithography for Samsung's 5LPE process node vs DUV lithography for Samsung's 8LPP process node.

So far, Qualcomm's current entry level Snapdragon SoC, the Snapdragon 4 Gen 1, is fabricated using TSMC's N6 process node. The only Snapdragon SoCs fabricated using Samsung's 4 nm** process node are the mid-range Snapdragon SoCs, specifically the Snapdragon 6 Gen 1 and the Snapdragon 7 Gen 1. Maybe that could change in the future, but who knows?

** → a marketing nomenclature used by all foundry companies

---

I'm seeing a lot of, I assume, high-end smartphones with 256 GB of UFS 3.1 storage and 12 GB of RAM. And the One Plus 9? I believe I saw for about $500. Would getting a deal with Samsung for a good number of components be able to get that in Drake for $449 in May at launch?

Probably not for the internal flash storage (probably 128 GB instead of 256 GB of UFS 2.1 or newer). But definitely a possibility for the RAM.

 

[Magic-Man]

Looks like you can access HBO Max through Hulu so I guess that's a solution for that. Agreed on Netflix though, it would've come out way sooner

HBO Max on Hulu isn't the full HBO Max, it just has a few shows and movies on it IIRC.

 

[Terrell]

Probably not for the internal flash storage (probably 128 GB instead of 256 GB of UFS 2.1 or newer). But definitely a possibility for the RAM.

I actually think you maybe have that backwards; of the 2, RAM is usually more expensive until you start crawling up past 256GB capacities for NAND. Especially right now.

I'm seeing a lot of, I assume, high-end smartphones with 256 GB of UFS 3.1 storage and 12 GB of RAM. And the One Plus 9? I believe I saw for about $500. Would getting a deal with Samsung for a good number of components be able to get that in Drake for $449 in May at launch?

When looking at smartphone prices to gauge part affordability, it's good to keep in mind that:

1. smartphones have to contend with:
   • parts not included in a console bill of materials, like multiple high-spec cameras, 5G antennae, GPS, wireless charging hardware and a host of other components
   • peculiarities in certain components not necessary for a gaming device that drive up costs, like screens with far more pixel density and cutouts for cameras/ear-speakers and smaller (thus higher-density) batteries
2. smartphones are sold at higher margins than consoles from day one; when looking at MSRP minus cost of goods sold (COGS), most smartphones are making between 40-60% revenue before deductions like marketing, retailer cuts/commissions and amortized R&D costs. For reference, using the same calculation method (x = MSRP - COGS), Switch was making ~14.3% revenue at launch.

Storage isn't typically one of the higher end part expenses in devices like this (it's not dirt cheap either, but it'll never be a major point of contention in COGS calculations like screens, SoCs, batteries and RAM are), but with the ideal conditions to get a favourable volume cost for it right now, yeah, I think 256GB of eUFS 3.1 is not out of the picture for a new hybrid game device at a $450 MSRP, possibly not even a $400 MSRP.

 

[Alovon11]

I actually think you maybe have that backwards; of the 2, RAM is usually more expensive until you start crawling up past 256GB capacities for NAND. Especially right now.

When looking at smartphone prices to gauge part affordability, it's good to keep in mind that:

1. smartphones have to contend with:
   • parts not included in a console bill of materials, like multiple high-spec cameras, 5G antennae, GPS, wireless charging hardware and a host of other components
   • peculiarities in certain components not necessary for a gaming device that drive up costs, like screens with far more pixel density and cutouts for cameras/ear-speakers and smaller (thus higher-density) batteries
2. smartphones are sold at higher margins than consoles from day one; when looking at MSRP minus cost of goods sold (COGS), most smartphones are making between 40-60% revenue before deductions like marketing, retailer cuts/commissions and amortized R&D costs. For reference, using the same calculation method (x = MSRP - COGS), Switch was making ~14.3% revenue at launch.

Storage isn't typically one of the higher end part expenses in devices like this (it's not dirt cheap either, but it'll never be a major point of contention in COGS calculations like screens, SoCs, batteries and RAM are), but with the ideal conditions to get a favourable volume cost for it right now, yeah, I think 256GB of eUFS 3.1 is not out of the picture for a new hybrid game device at a $450 MSRP, possibly not even a $400 MSRP.

Yeah, and because of the economy of scale for a console, and how it would be the only (at least for a bit), SKU using that Memory from Nintendo, that would help bring costs down further.

 

[Hermii]

Somehow everybody concluded that TX1+ is using 16nm.

Thats definitely not how I remember it. Once we knew the clock speed, the game was up.

Edit: I read your post wrong, didn’t see the +.

 

[Hermii]

Without knowing anything about the context of that chart I really don't think it can inform us much. Is the tool using the 4.2W number to estimate the entire GPU power draw? Or maybe just part of it. Does Drake have some major power efficiency improvements that Orin does not? Could it have been an old out of date test done when the GPU had a different configuration with different power requirements?

There's too much we don't know to draw solid conclusions.

I agree, making conclusions based on that chart is shaky. But making conclusions based on the clocks and power draw of the final product is not.

Drake does probably have efficiency gains over Orin and other ampere cards, but there is no way better clock gating and cutting some features doubles the performance per watt of the chip.

And if it somehow miraculously does, who cares what node it is. We are still getting more than we hoped for.

 

[mariodk18]

So...Do you think Nintendo and Nvidia talked to each other between 2017 and 2019 and tried to come up with the most powerful late 2022 hardware possible for roughly ~$400? Because that kind of seems like what we might end up getting. Unless I'm totally off base of course.

 

[Dakhil]

I actually think you maybe have that backwards; of the 2, RAM is usually more expensive until you start crawling up past 256GB capacities for NAND. Especially right now.

I assume you meant to say internal flash storage (or as smartphone manufacturers erroneously say, ROM)?

 

[ReddDreadtheLead]

So...Do you think Nintendo and Nvidia talked to each other between 2017 and 2019 and tried to come up with the most powerful late 2022 hardware possible for roughly ~$400? Because that kind of seems like what we might end up getting. Unless I'm totally off base of course.

No, what I think happened is that Nintendo took the feedback of their internal development studios and external development studios, juxtaposed what they were asking for and presented it to nvidia. Nvidia were hired to deliver a product that fits the meds that Nintendo outlined in the data they collected from developer feedback.

Nvidia provided a prototype to which they informed Nintendo of what they can work with, Nintendo is willing to pay for it as making a chip for them isn’t that expensive for their needs.

Nvidia “easily” delivers something that suits Nintendo’s needs and includes features that Nintendo would utilize and make sense in a game console. Also due to what Nintendo was willing to pay.


Not only that but they had to negotiate with nvidia a price for the chip and a plan for the chip long term and how it’ll depreciate.

They look at what they need for the chip as well long term like RAM and Storage, and other components related to it like the game cards, etc.



If it is the most powerful handheld, that’s only due to the circumstances resulting in that, but it wasn’t really their goal per se. Just a byproduct of the circumstances.

 

[Terrell]

I assume you meant to say internal flash storage (or as smartphone manufacturers erroneously say, ROM)?

No, I meant it, Random Access Memory is, as a general rule, a more expensive part in most electronics than NAND flash storage until you begin going past a certain threshold, and with the tech developments in play and the current NAND pricing, that gap is getting wider.

 

[Jersh]

Does Drake have some major power efficiency improvements that Orin does not?

Z0m3le said what I was thinking here:

if the GPU is twice as efficient as Orin at the same configuration and clock, it really doesn't matter what node we think it is, the end result is that it performs like a 5nm Orin GPU. Shrug you want to assume it's 8nm, there is only so much that actually matters.

The primary reason we care about this at all is efficiency. Sure, it wouldn't be across the board (the CPU wouldn't benefit, and potentially some components of the GPU like RT cores), but at the end of the day it doesn't matter if the extra efficiency comes from a die shrink or an architectural iteration, they achieve the same thing. If they improved Drake so much on 8nm that it preforms like Orin would on 5nm (or 6N, or whatever it corresponds to) then it may as well be on that node for the subject of our speculative ballparking (though I'll also note that this feels rather unlikely).

To me, none of the other ramifications matter at this stage. The chip will physically fit in the unit or the device wouldn't exist. We can use node info to try to pinpoint Nintendo's price target, but there's so many other variables (screen? storage? materials? memory configuration?) that it's a crapshoot anyway.

I agree with the take that all of the info in question is, well, in question, if it's info at all. But I do think the notion that it's too wildly off of where 8nm Ampere is right now to have the same performance characteristics is on the right track.

 

[Codename_Steam]

It's unknown actually. Drake's OFA engine is more advanced than Ampere, it's unknown how it stacks up against Ada's, but it is for auto AI, and is the more extreme application, we just don't know what that means exactly, however Frame Generation can technically be done slower on Ampere, and when we are talking about 30fps of frame generations, that is likely possible, given that DLSS 3.0 on Ada can handle 100fps+.

Does Drake feature any Lovelace enchantments? (Like Mixed-Precision on Tegra X1)

 

[ReddDreadtheLead]

Does Drake feature any Lovelace enchantments? (Like Mixed-Precision on Tegra X1)

No, it’s a bonafide boring ampere. Nothing interesting about it.

 

[Concernt]

No, it’s a bonafide boring ampere. Nothing interesting about it.

It has a new CUDA version, but that could just be some changes to efficiency or for Maxwell compatibility.

 

[ReddDreadtheLead]

It has a new CUDA version, but that could just be some changes to efficiency or for Maxwell compatibility.

That doesn’t really matter in this case, it’s basic ampere.

 

[ILikeFeet]

No, it’s a bonafide boring ampere. Nothing interesting about it.

How dare you, does this look uninteresting to you?

---

UE5.2 features skinned mesh support for nanite

 

[ReddDreadtheLead]

How dare you, does this look uninteresting to you?

Vanilla ice cream has more flavor and character

 

[ReddDreadtheLead]

But I do think the notion that it's too wildly off of where 8nm Ampere is right now to have the same performance characteristics is on the right track.

What do you mean?

 

[kvetcha]

What do you mean?

I believe that @Jersh is agreeing that the power and frequency estimates provided in the tester, as well as the estimated size of a Drake-like SOC, hints that Drake is something other than Samsung 8nm, since Ampere's known performance characteristics on that node are drastically misaligned

 

[Goodtwin]

running Nvidia's estimation tools for Orin with everything set to low and the gpu off, the system power draw is estimated at 8.5w, with the GPU set to 4TPC and 624mhz, it draws 14.5w, that is 6w for 1024 cuda cores on 8nm at 624MHz, Drake is being estimated to draw just 4.2w at 660MHz on 6TPC or 1536 cuda cores... Without a single doubt, Drake can't be 8nm if the estimation is accurate, and with a disparity this great, it's very very unlikely it wasn't caught.

I disagree whole heartedly. There is no way you could estimate Ampere with 12SM to draw 4.2w on 8nm at 660MHz, it's double that.

The question is how much of a node reduction does it typically take to get the improved efficiency we are seeing here. I agree with some others who think the 4N process would likely be too expensive for Nintendo. Does moving to 6N get the density and power efficiency improvements needed here.

On a separate note, with Nintendo likely keeping the OG Switch around for a few more years, does Nintendo feel like a $499 launch price is on the table? Nintendo knows that it's Switch audience is largely made up of young adults and may feel comfortable with selling a high end Switch for a year or two before dropping the price.

 

[ILikeFeet]

The question is how much of a node reduction does it typically take to get the improved efficiency we are seeing here. I agree with some others who think the 4N process would likely be too expensive for Nintendo. Does moving to 6N get the density and power efficiency improvements needed here.

On a separate note, with Nintendo likely keeping the OG Switch around for a few more years, does Nintendo feel like a $499 launch price is on the table? Nintendo knows that it's Switch audience is largely made up of young adults and may feel comfortable with selling a high end Switch for a year or two before dropping the price.

honestly, not much. you have to remember that Samsung 8nm is an enhanced 10nm, which is from 2017. it's old as shit. any newer node would be better, like Samsung 4nm (since their 5nm is their 10nm but with EUV layers). moving to TSMC would be a big jump since their 6nm is a legitimate jump and will be a long-lasting node. TSMC 5nm is what Lovelace is on (as 4N) and is the most efficient gpu in the desktop market right now

as for a $500 price, I doubt it's on the table. I think $450 is the highest they'll go, but who knows

 

[Deleted member 887]

So...Do you think Nintendo and Nvidia talked to each other between 2017 and 2019 and tried to come up with the most powerful late 2022 hardware possible for roughly ~$400? Because that kind of seems like what we might end up getting. Unless I'm totally off base of course.

This is basically what happened. There are strong indicators in the leak about Nintendo/Nvidia's thought process, but just going by public data...

DLSS 2.0 and Ampere were over a year out from release when Nintendo committed to using them as the base of their next hardware. I don't know how far along Lovelace was at the time, but it was in development already. The A78 was also a year out from release when Nintendo committed to it as their processor core, and the A710 design was absolutely available for review and to begin integration

Nintendo made a strong bet on bleeding edge tech 3 years ago, betting on it maturing in time for launch. That bet has panned out in a number of ways. On the processor side, the A78 design has proven itself with some of its successors either just being lightly retooled, underperforming, or both. DLSS has rapidly improved as integrations have spread. Ampere and Lovelace are very similar architectures, and Nintendo has almost been able to pick and chose what it wants from the two arches, while dodging some of the capacity problems that have plagued the industry.

It's entirely possible that inflation will force Nintendo to either increase the price or sell at a loss, but they're in their best position ever to do one of those things, not to defend their market position, but to grow it.

 

[ILikeFeet]

This is basically what happened. There are strong indicators in the leak about Nintendo/Nvidia's thought process, but just going by public data...

DLSS 2.0 and Ampere were over a year out from release when Nintendo committed to using them as the base of their next hardware. I don't know how far along Lovelace was at the time, but it was in development already. The A78 was also a year out from release when Nintendo committed to it as their processor core, and the A710 design was absolutely available for review and to begin integration

Nintendo made a strong bet on bleeding edge tech 3 years ago, betting on it maturing in time for launch. That bet has panned out in a number of ways. On the processor side, the A78 design has proven itself with some of its successors either just being lightly retooled, underperforming, or both. DLSS has rapidly improved as integrations have spread. Ampere and Lovelace are very similar architectures, and Nintendo has almost been able to pick and chose what it wants from the two arches, while dodging some of the capacity problems that have plagued the industry.

It's entirely possible that inflation will force Nintendo to either increase the price or sell at a loss, but they're in their best position ever to do one of those things, not to defend their market position, but to grow it.

I honestly wouldn't be shocked if, in 2019, they made bets on Samsung having an improved fab. that didn't work out, but at the same time, Samsung is getting better

 

[Deleted member 1324]

I’m not sure what AMD can offer for SONY/MICROSOFT for a console targeting 2024 as revisions.

Twice the GPU compute for what they paid for their launch consoles in 2020.

Their consoles are already struggling to get anywhere near 4k even with reconstruction. Some games are also struggling to hit 30fps with ray tracing due to being GPU bound at 1440p/1080p.

If you double the GPU compute lots of games would be far nearer native 4k and the games that use RT would be a locked 30/60fps depending on the game instead of dropping into the 20/50’s like they do now.

I don’t expect the upgraded PS5 and Series X until late 2024 though not late 2023 like some people do.

 

[Intoxicate]

Twice the GPU compute for what they paid for their launch consoles in 2020.

Their consoles are already struggling to get anywhere near 4k even with reconstruction. Some games are also struggling to hit 30fps with ray tracing due to being GPU bound at 1440p/1080p.

If you double the GPU compute lots of games would be far nearer native 4k and the games that use RT would be a locked 30/60fps depending on the game instead of dropping into the 20/50’s like they do now.

I don’t expect the upgraded PS5 and Series X until late 2024 though not late 2023 like some people do.

… and then the PS5 Pro will be launched, promising 4K for all games that struggle to get there. Wait, I thought that’s why I upgraded my PS4 to a Pro last time and still the Pro didn’t deliver. Anyway the motivation to display 4K is nill for me. I don’t care for movies and even less for games. But it’s my fault, I just don’t want to move closer to my TV.

 

[JoshuaJSlone]

PS4 Pro and Xbox One X didn't really put up big numbers, did they? And there's no worthwhile number of screens beyond 4K to get people to upgrade for. If they do mid-gen revisions, their target will need to be people incredibly dissatisfied with the most strenuous games being 4K30.

 

[Deleted member 887]

I'm going to again preface this by saying I have a shallow understanding of the things I'm seeing at this point. But this certainly seems relevant to the thread.

NVN v1 source has a constants file with the following:

// Number of warps per SM on TX1 hardware
#define __NVN_NUM_WARPS_PER_SM_TX1                  128

// Number of SMs on TX1 hardware
#define __NVN_NUM_SMS_TX1                           2

NVN v2 source replaces them with this:

// Number of warps per SM on ga10f
#define __NVN_NUM_WARPS_PER_SM_GA10F                48

// Number of SMs on on ga10f
#define __NVN_NUM_SMS_GA10F                         12

amount of register memory and the capability of the thread scheduler and stuff like that. The actual amount of warps you can run concurrently on an SM isn't always exactly that max, either, depending on how you're using them. It does indicate that there are more low-level differences between desktop Ampere and T239 than we're aware of, but my guess is that it's probably shared with Orin, and may be one of a number of low-level optimisations around power consumption.

I realize I am replying to old messages here, but this damn thing is gonna get announced soon, and I want to make sure I get it right before the game ends

Warps = Compute Shaders = Pixel Shaders. Drake's SMs have the same number of partitions, registers, register memory, and CUDA cores as desktop SM, but for some reason limits the number of warps available to 3/4 of the capacity of desktop. This is not VTG - vertex/tessellation/geometry - shaders. Usually this is half the number of warps, but it's not for Drake.

My understanding is that pixel shaders have become far less common than vertex shaders in modern games. Because everything else is the same, this is either a software limitation, or they've pulled scheduling hardware for compute out of the SM. I can't imagine this saves power, but it might save die size. RT runs at the compute stage of the pipeline, but I don't know if those shaders are scheduled the same way.

Does anyone have strong knowledge of how the lack of compute shaders relative to VTG shaders might impact perf?

 

[ILikeFeet]

the wider public doesn't give a shit about native 4K or even stable frame rates, pro systems don't have much of a selling point to justify the extra cost

 

[SiG]

the wider public doesn't give a shit about native 4K or even stable frame rates, pro systems don't have much of a selling point to justify the extra cost

That is, assuming they do not have any other gimmicks...

 

[Deleted member 887]

PS4 Pro and Xbox One X didn't really put up big numbers, did they? And there's no worthwhile number of screens beyond 4K to get people to upgrade for. If they do mid-gen revisions, their target will need to be people incredibly dissatisfied with the most strenuous games being 4K30.

No one knows the exact sales, because they blend the sales numbers with the base consoles. They probably did pretty well, and continue to do so. PS 5 Pro and Xbox Series XXX probably aren't coming though.

Microsoft has been explicit that the reason they made the Series S was that they didn't expect the die shrinks that would make a slim console viable - and an upclocked more powerful version - to be there. And they've mostly been proven right. They've already got two, wildly different SKUs that they require developers to support with simultaneous launches, they're not going to add a third. And how would they justify it to customers without the 4k push that got the pro consoles out there in the first place?

Sony can't get developers to really use the hardware as it stands, and they've got more generational exclusives than Microsoft. Shoving more hardware in there isn't going to help matters, and it would cost a butt. Sony is going to keep finding ways to make them cheaper to build and to ship, but I doubt they're going to shove more power in there. There isn't a technical driver, there isn't a market driver.

 

[ILikeFeet]

That is, assuming they do not have any other gimmicks...

that would be against Sony and MS's MO of "the same but more power". other than VR, there's not much avenue I can think of that isn't something both has already attempted and dropped

 

[Deleted member 1324]

if the GPU is twice as efficient as Orin at the same configuration and clock, it really doesn't matter what node we think it is, the end result is that it performs like a 5nm Orin GPU. Shrug you want to assume it's 8nm, there is only so much that actually matters.

You still predicting around 2TFLOP portable and 4TFLOP Docked for Drake?

 

[Jersh]

What do you mean?

Once again I've been beat to it:

I believe that @Jersh is agreeing that the power and frequency estimates provided in the tester, as well as the estimated size of a Drake-like SOC, hints that Drake is something other than Samsung 8nm, since Ampere's known performance characteristics on that node are drastically misaligned

Now sure, maybe those numbers aren't even supposed to represent the full GPU power consumption, or something else is at play that we're missing. I'm not claiming anything as a certainty. But it's off of Nvidia's own estimation tools/data by so much that I feel the simplest and most likely explanation is that it's a notably more efficient chip.