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

[Skittzo]

So with these latest revelations, are we still in the belief that NG Nintendo is around PS4+ levels of 'power'?

Effectively (when docked) it'll be anywhere around PS4Pro to around Series S power, yeah.

Undocked it's basically a PS4 with major architectural improvements

 

[Deleted member 887]

One thing I don't understand is how is AMD so so far behind? When did this start?

I'm gonna actually try and answer this question, so buckle up

TL;DR: Nvidia bet on features, AMD bet on power. Features won in the market, which left AMD spending their extra power to simulate those features, leaving them with less power to go around, and features that aren't as good.

Every big bump in resolution roughly doubles how much detail a human eye can pick up, but roughly quadruples the number of pixels. And because it quadruples the pixels, it quadruples the amount of power it takes to put that stuff on screen. If you think about that for more than a minute, the problem becomes obvious - this shit can't go on forever.

And that resolution leap doesn't include making this pixels prettier. Not just advanced details but advanced effects, like higher quality lighting, reflections, etcetera. So you need to quadruple performance just to stand still. You need to do better than that to advance.

In every field except the GPU, those advances in performance have become extremely difficult. At some point, making CPU's faster got really hard, which is why they put more and more cores in every generation. GPUs happen to scale very well with adding more cores, so GPUs have dodged the wall that other system's have been hitting. But that won't last forever.

Both Nvidia and AMD clearly saw this writing on the wall. Neither of them (or Intel, in fact, but that's a tangent right now) misunderstood the problem. What happened next is that they tried two very different solutions.

AMD is a secondary player in the desktop space, with a lot of their core consumers being budget players. They absolutely dominate consoles, and have for the last 2 decades. They're a strong player in the data center, and they have a CPU product that dominates the industry, and is based on a technology called "chiplets" where they can mix and match parts from different foundries. That lets them rapidly customize products, while also manufacturing performance critical chunks of a chip with the most advanced but expensive tech, and less performance critical chunks on cheaper tech.

AMD's strategy was this - keep pursuing that classic gen-on-gen power, by iterating on their core design. Keep it backwards compatible for their console customers. Keep their data center and consumer segments different, but invest heavily in bringing their chiplet tech to GPUs. That will allow them to very quickly adapt products to the market without having to design new hardware from scratch each time, while also keeping costs down.

AMD saw the wall coming, laid down the gauntlet and said, fuck it, we're going to bust straight through that thing. It was smart and aggressive.

That... is not what Nvidia did. Nvidia decided that the only winning move was not to play. Nvidia decided that instead of pursuing more and more power they would pursue features. Nvidia added Ray Tracing, which doesn't make More Pixels, but does make Prettier Pixels. And because it's a relatively new tech for the consumer space, there is a much much longer road of innovation ahead of them, betting they can deliver huge leaps on the RT side while the traditional rasterization side slows down.

They didn't pursue chiplets, instead deciding to just make their datacenter designs and their consumer designs the same to reduce design costs. That meant putting AI hardware on consumer products, AI being another feature where huge leaps are still possible. And that meant finding a use for that AI hardware in the first place.
Which lead to AI assisted upscaling, and DLSS in the first place.

Early on, it looked like AMD had pulled it off. Low RT powered device couldn't deliver much, and few games took advantage of it without Nvidia throwing money at it. AMD figured out how to add basic RT to their hardware with minimal modification, instead of the huge investment Nvidia made. At "traditional" rendering, AMD was delivering better performance, and AI upscaling wasn't just bad, it invented new kinds of bad no one had ever seen before. Bad upscaling would miss detail, and just create blurrier images. DLSS 1.0 was instead finding detail that didn't exist and looked wrong adding bizarre details to images that made no sense. And it was expensive requiring a supercomputer to train an custom AI model for each game that wanted to use it.

Then 2020 happened. Control was out, and people started to see RT could really do, and because AMD had at least minimal support, RT modes started to become common in games, which of course ran better on Nvidia. And then came DLSS 2.0 which was a generic solution, easy to implement, that didn't have DLSS 1.0's problems and actually delivered on the AI promise - and that only worked on Nvidia cards.

DLSS 2.0 and RT actually don't interact super well with each other, but Nvidia very smartly figured out how to make them seem like they did. Instead of selling DLSS 2.0 as a way to make resolutions higher, it sold it as a way to make frame rates higher - as a tech that could recover the performance "lost" by enabling RT. So even though the two technologies actually fight each other a little bit under the hood, Nvidia managed to find a way to make them seem tied at the hip.

So Nvidia had features, but arguably, AMD had power and cost. But it was about to get worse for AMD. Even with the power of advanced GPUs, developers were having trouble pushing all those damn pixels with 4k everything, and so they started to use temporal upscaling - the same class of tech as DLSS 2 - everywhere. AMD released a best-in-class upscaler (FSR 2) which delivered similar results to DLSS without the tensor cores.

On paper that's great - it is great! - but it meant that all that extra power AMD had bet on was being used to replicate Nvidia features. Game X might run better on AMD than Nvidia out the gate, but then enable DLSS 2.0 and NVidia runs much better. AMD brings out FSR2 to match, but FSR2 itself eats the extra power that gave AMD the advantage in the first place. And it doesn't look quite as good without that AI to help.

Then came the RTX 40, and jaws dropped. Prices were awful because the advanced foundry nodes that GPUs had been rushing to for decades were getting more and more expensive, and without chiplets, Nvidia was carrying that cost on every single square millimeter on their new chip. This was exactly what AMD had expected, and why they invested in chiplet designs.

Months later and the RX 7000 series was revealed and prices... were just as bad. AMD had pulled it off, they had managed to build a chiplet GPU. But it turned out to be a very different problem than a chiplet CPU, and because of that, very little of the GPU could be built on a cheaper node, and thus, very little cost savings on this first version of the design.

And there were other problems with RX 7000 as well. AMD updated their cores to have some of Nvidia's advantages that made DLSS/RT fast - like dual issue compute, and accelerated matrix instructions. But the commitment to backwards compatibility was showing it's age, with lots of complexity in the front end making utilization of these features fall way short of their theoretical max.

But Nvidia does have chickens coming home to roost. AMD might not have nailed it, but they weren't wrong. Chiplets are the future, and Nvidia has to get there. Nvidia didn't skip the chiplet investment, they just delayed it. And in this time of surging AI products, AMD's chiplet design is paying off. They're able to put together custom data center products that combine several of their technologies extremely quickly.

When it comes to backwards compatibility, Nvidia and Nintendo have likely invested huge quantities of money to make it happen, and will probably have to do so again in 5 years. If their BC is emulation driven, then Nvidia is developing the software that make it possible for Nintendo to go to a different vendor in the future. AMD has gotten said BC nearly for free, and has locked in the other two console makers likely for a couple more generations.

AMD is also innovating, with the recent previews of Frame Gen technology that works in legacy games without patches. That's potentially a huge win for the PS6/Next Box, allowing 120fps modes for everything but also potentially a major win for those handheld PCs whose value proposition is often around being able to run last gen games in your hand.

And AMD is likely to dominant in the handheld PC space not just because they have top-tier PC CPU, but because again, this is a place where their chiplet tech has huge potential to pay off, with AMD able to deliver customized APUs extremely quickly, and with a low enough design cost that a customized APU is affordable even for products that don't sell millions of units.

It remains to be seen if AMD can deliver on the potential of chiplets, and can catch up on the feature space. But it also remains to be seen how much further Nvidia can take DLSS, with Frame Gen and Ray Reconstruction being the obvious evolutions of the tech. Nvidia has got a big roadblock with their move to chiplets, but AMD actually already has top-class machine learning hardware in their server offerings, and could catch up rapidly if they decide to go that path.

May you live in interesting times!

 

[Goodtwin]

How much would Pratt even cost for something like this ?
There's just no way Nintendo would hire one of the top Hollywood actors for a game.

Voice acting does not take a lot of time. Jennifer Hale did all the voice acting for Bayonetta 3 in a few days. So yes, Chris Pratt would probably cost more than the typical voice actor, but for a videogame that will go on to sell tens of millions of copies, even a million dollar expense to get Pratt would be within the budget.

 

[yusufied]

What if we get something at the next Nintendo direct like "join us this fall for the reveal of our next console" that would be so hype

Yeah fr just tell me it exists and stop tip toeing around the subject. I know they're worried about the holiday sales but are they seriously worried about the laggards in the market putting off getting a Switch for the NG? We're approaching 7 years since launch at this point.

 

[Thraktor]

One thing I don't understand is how is AMD so so far behind? When did this start? I know Nintendo used to be on AMD/ATI then moved to Nvidia with Switch and apparently were even considering moving with the 3DS. By pure chance they set up their future success beyond the Switch itself.

The advantages of being with Nvidia are going to help the Switch 2 match or maybe even supercede the Xbox Series S when it comes to docked resolution. DLSS would help the PS5 and Xbox Series X a lot but they're struggling with FSR instead. Do you guys think FSR will ever catch up? It's not even just FSR, Nvidia's RT is much much better too.

It's honestly kind of embarrassing for owners of anything that isn't Nvidia based LOL

The real issue is that AMD spent a long time barely holding on financially. The Bulldozer CPU architecture (and its successors) were disastrous for the company, and they lost almost all of their market share in the (high margin) server sector, and could only hang on to a small part the (low margin) entry level of the PC space. The GPU business wasn't doing completely terribly at the time, but CPU sales have always been the core of AMD's revenue stream, so they were struggling to make a profit as a company.

What do you do when you're struggling to make a profit and there's no quick way to increase sales? Cut back on any expenses that don't have a direct path to profitability. That means dropping exploratory R&D on long-term technologies like hardware ray tracing and AI. It also meant cancelling their planned "K12" ARM CPU core to focus on Zen instead (a sensible move in retrospect). AMD spent some very lean years working on a slim R&D budget that could only really justify straight-forward technological advances, meaning new CPU and GPU architectures that do the same thing, but faster. They couldn't justify the spend on something like hardware ray tracing that's not guaranteed to pay off.

Even their biggest innovation of the past few years was largely motivated by minimising R&D cost and risk. The reason they started using chiplets when Intel was still entirely focused on monolithic chips is that designing and taping out chips on leading-edge nodes costs a lot of money, money which AMD didn't have at the time. By moving to a chiplet approach, AMD could tape out just one chip on a leading-edge node, plus one I/O die on an older, cheaper node, and cover everything from entry-level desktops to 64-core servers. They were looking to do with one die what Intel were doing with 5 or 6, which they did quite successfully.

It was only really in 2019 that AMD's financials started to turn around. In 2019, AMD's revenue was $6.73 billion, by 2022 that had gone up to $23.6 billion. The payoff of the Zen architecture was slow, because although home PC builders adopted Ryzen pretty quickly, it took time to convince OEMs/server/HPC customers/etc. to actually consider AMD chips again. They're now in a much better place financially, so they can start investing in more of that fundamental R&D again, but it takes a long time to start up that kind of research. You have to hire experts in the area, you have to work out all the low-level fundamental details before you can start designing hardware, and even when you've designed the hardware it will be a year or two before it's in anyone's hands.

In terms of AI acceleration, I'd say AMD is actually in a pretty good place from a hardware point of view. They've been shipping HPC chips with matrix cores (basically the same thing as tensor cores) for several generations now, although initially they were more focussed on higher-precision work for HPC applications (ie really good FP64 performance). For pure AI use-cases, they seem to have made significant improvements, and it's quite possible that AMD's new MI300X flat-out outperforms Nvidia's H100 with equivalently optimised software.

Therein lies the problem, though, software. Nvidia spent a lot of time and money not just developing AI hardware, but developing the software packages and making sure they're widely used. Almost any AI software stack you'll find out there has far better support for Nvidia hardware than AMD (if it runs on AMD hardware at all), and that's a big hill AMD have to climb. One thing that's going for them is that, with the spike in demand for AI hardware, the industry is suddenly becoming a lot more interested in making sure they have other hardware options than just Nvidia. Nvidia is charging through the nose for H100s, which is making them a lot of money, but creates a big incentive for AI software that works well across a range of hardware, which AMD will look to capitalise on.

On the consumer side, AMD have added matrix cores to their GPUs starting with the RDNA3 architecture, but as yet haven't been using them in games. They didn't really publish performance figures on them, but this blog post states they can get 512 ops/clock/CU for FP16 and BF16. That would come to about 122 Tflops on the RX 7900XTX, which is about the same as an RTX 4070, which is to say more than enough for something like DLSS. Again, there's a lack of software there on AMD's side, which comes from them not having the head-start that Nvidia had. I'd wager we'll see a AI-based version of FSR at some point in the next few years, though.

On the ray tracing side of things, it's a trickier problem to solve. They added hardware triangle intersection testing pretty quickly, but it's a relatively simple bit of circuitry, so it was an easy win for them. I have no doubt they're working on hardware accelerated BVH traversal, but it's not as easy a problem to solve in hardware as triangle intersection testing, and I'm betting Nvidia were working on it for a long time before actually launching Turing. I'd expect to see it in the next generation or two of AMD GPUs, at which point it will probably close the gap significantly.

TLDR: AMD had no money for ages so couldn't afford R&D on things like RT and AI. They now have money, so can fund this R&D, but it takes time.

 

[Deleted member 887]

So with these latest revelations, are we still in the belief that NG Nintendo is around PS4+ levels of 'power'?

For me, yeah. There aren't a lot of details, but what's leaked out matches what I thought would be possible with that combination of performance and Nvidia hardware. So I feel like the high-level estimates are true

 

[fwd-bwd]

so the three options are, dealing with incorrect lighting, turn all the lighting off so the game now looks flat like Detective Pikachu 2, or making sure you're lighting setup works in both handheld and docked mode

Not replying to you in particular, just thinking out loud. I’ve been wondering about the remaster of Luigi’s Mansion Dark Moon. The title seems a random choice until we consider its potential for lighting experiments. It could be a test bed of sorts for Nintendo’s team to experiment and establish cross-gen lighting solutions that would look good on both the NG (with RT) and Switch 1 (without). Being a remaster, they can get the game up and running in a timely manner at a relatively low cost, and focus on developing best practices for graceful degradation between two gens with vastly different power levels.

It would not surprise me if Chris Pratt is the voice of Mario for the next game and the animation is very similar.

I’d like to think that Nintendo would rather spend money on game development (e.g., physics engine for 3D Mario) than celebrity VOs, but the company is very capable of surprising people so I don’t really know.

 

[SoldierDelta]

For me, yeah. There aren't a lot of details, but what's leaked out matches what I thought would be possible with that combination of performance and Nvidia hardware. So I feel like the high-level estimates are true

For what it's worth (with what we know of the specs for the Sw2), would Hero's claim about FF7R's graphics (not resolution) being on par with the PS5 have any weight with any of the new information that we got recently?

I know the claims from Hero aren't exactly well accepted by a lot of people here, but I'm mostly asking as a hypothetical.

 

[Mr.Gamerson]

Japanese devs are going to eat on Switch 2 at least.

I hope for the more demanding games we get

540p to 1080p DLSS handheld
720 to 1440p DLSS Docked

Which would be great overall. I'm really interested to see how well 360p to 1080p Dlss and 540p to1440p dlss will translate for handheld and docked modes.

 

[mjayer]

You guys do realize the Mario movie is one of the most successful animated movies of all time, right?

 

[Deleted member 887]

I’d like to think that Nintendo would rather spend money on game development (e.g., physics engine for 3D Mario) than celebrity VOs, but the company is very capable of surprising people so I don’t really know.

Nintendo underpays even for their unknown VOs. The video game industry as a whole aggressively pursues non-Union talent, but Nintendo especially is known for pursuing amateur talent. Charles is a little unusual for being a professional actor, probably because he didn't start working on Mario for video games at all, but for their trade shows (which are non-union gigs) and only got kept on for the video games because Miyamoto liked him personally.

This seems to be changing a bit, with "proper" voice actors doing more and more work. But I can't imagine them going with Pratt, not because of the cost, but the scheduling. No way would Pratt's agent let him sell perpetuity rights like Martinet did, meaning voice clips couldn't be used in future games, so they'd have to find time to schedule voice recording sessions around a Hollywood actor's schedule, on a game development timeline. Nightmare

 

[lemonfresh]

You guys do realize the Mario movie is one of the most successful animated movies of all time, right?

I just checked and yeah youre right damn

 

[Deleted member 887]

You guys do realize the Mario movie is one of the most successful animated movies of all time, right?

Why is that relevant?

 

[Serif]

For what it's worth (with what we know of the specs for the Sw2), would Hero's claim about FF7R's graphics (not resolution) being on par with the PS5 have any weight with any of the new information that we got recently?

FF7IIR on Switch 2 looking like Intergrade is plausible as the differences include resolution, lighting and texture improvements.

From our understanding of Switch 2 as a PS4+, it could render a 1080p image with bells and whistles and AI upscale to 1440p/2160p, making a sharper image comparable to the PS5 version.

So this specific claim by Hero isn't out of line with what we anticipate.

 

[Danny]

Yeah i mean for a video game company that rarely even use voice acting in their games it seems unrealistic that they would then hire a bigger star than even Playstation studios hire to perform as a voice actor on their games. I mean 1 playstation game usually has the same amount of voice acting as 100 Nintendo games combined.

 

[Concernt]

You guys do realize the Mario movie is one of the most successful animated movies of all time, right?

So?

That doesn't mean we want that ugly ass Mario in a GAME. He was redesigned to work with a film, but games aren't film, and so they don't need, and shouldn't use, the same designs. Why give up that iconic face to pander to movie fans that will buy a Mario game anyway? They don't care that he looks a little different. It's just a Mario game

 

[Skittzo]

Good lord let it go, they're not using Chris Pratt as Mario in games from here on out.

 

[Thraktor]

You guys do realize the Mario movie is one of the most successful animated movies of all time, right?

Everyone knows that, but I think people are overstating its importance to Nintendo. In Nintendo's most recent quarterly report, the revenue from mobile/IP/etc. (which includes revenues from the Mario movie) increased to 31.8 billion yen from 10.9 billion yen. Let's assume the full 21 billion yen increase was from Mario movie revenues. That's about $142 million at current exchange rates, from the one point whatever billion of box office revenue.

Mario Odyssey has sold 26.44 million units. Assuming Nintendo sees on average $40 from each copy (probably an underestimate with growing digital ratios, but it's a round number), Nintendo has pulled in over a billion dollars from Mario Odyssey alone, let alone any other Mario games on the system (MK8 is probably over $2B).

The Mario movie is a great way for Nintendo to grow their brand and bring in extra licensing revenue, but from a financial point of view it's on about the same scale as one of their lowest-performing video games. Mario games are exceptionally profitable for Nintendo, and I don't see them messing with a formula that's worked so well for them just to better fit in with a series of movies that's bringing in a tiny fraction of the revenue.

 

[Bigdog]

I'm gonna actually try and answer this question, so buckle up

TL;DR: Nvidia bet on features, AMD bet on power. Features won in the market, which left AMD spending their extra power to simulate those features, leaving them with less power to go around, and features that aren't as good.

Every big bump in resolution roughly doubles how much detail a human eye can pick up, but roughly quadruples the number of pixels. And because it quadruples the pixels, it quadruples the amount of power it takes to put that stuff on screen. If you think about that for more than a minute, the problem becomes obvious - this shit can't go on forever.

And that resolution leap doesn't include making this pixels prettier. Not just advanced details but advanced effects, like higher quality lighting, reflections, etcetera. So you need to quadruple performance just to stand still. You need to do better than that to advance.

In every field except the GPU, those advances in performance have become extremely difficult. At some point, making CPU's faster got really hard, which is why they put more and more cores in every generation. GPUs happen to scale very well with adding more cores, so GPUs have dodged the wall that other system's have been hitting. But that won't last forever.

Both Nvidia and AMD clearly saw this writing on the wall. Neither of them (or Intel, in fact, but that's a tangent right now) misunderstood the problem. What happened next is that they tried two very different solutions.

AMD is a secondary player in the desktop space, with a lot of their core consumers being budget players. They absolutely dominate consoles, and have for the last 2 decades. They're a strong player in the data center, and they have a CPU product that dominates the industry, and is based on a technology called "chiplets" where they can mix and match parts from different foundries. That lets them rapidly customize products, while also manufacturing performance critical chunks of a chip with the most advanced but expensive tech, and less performance critical chunks on cheaper tech.

AMD's strategy was this - keep pursuing that classic gen-on-gen power, by iterating on their core design. Keep it backwards compatible for their console customers. Keep their data center and consumer segments different, but invest heavily in bringing their chiplet tech to GPUs. That will allow them to very quickly adapt products to the market without having to design new hardware from scratch each time, while also keeping costs down.

AMD saw the wall coming, laid down the gauntlet and said, fuck it, we're going to bust straight through that thing. It was smart and aggressive.

That... is not what Nvidia did. Nvidia decided that the only winning move was not to play. Nvidia decided that instead of pursuing more and more power they would pursue features. Nvidia added Ray Tracing, which doesn't make More Pixels, but does make Prettier Pixels. And because it's a relatively new tech for the consumer space, there is a much much longer road of innovation ahead of them, betting they can deliver huge leaps on the RT side while the traditional rasterization side slows down.

They didn't pursue chiplets, instead deciding to just make their datacenter designs and their consumer designs the same to reduce design costs. That meant putting AI hardware on consumer products, AI being another feature where huge leaps are still possible. And that meant finding a use for that AI hardware in the first place.
Which lead to AI assisted upscaling, and DLSS in the first place.

Early on, it looked like AMD had pulled it off. Low RT powered device couldn't deliver much, and few games took advantage of it without Nvidia throwing money at it. AMD figured out how to add basic RT to their hardware with minimal modification, instead of the huge investment Nvidia made. At "traditional" rendering, AMD was delivering better performance, and AI upscaling wasn't just bad, it invented new kinds of bad no one had ever seen before. Bad upscaling would miss detail, and just create blurrier images. DLSS 1.0 was instead finding detail that didn't exist and looked wrong adding bizarre details to images that made no sense. And it was expensive requiring a supercomputer to train an custom AI model for each game that wanted to use it.

Then 2020 happened. Control was out, and people started to see RT could really do, and because AMD had at least minimal support, RT modes started to become common in games, which of course ran better on Nvidia. And then came DLSS 2.0 which was a generic solution, easy to implement, that didn't have DLSS 1.0's problems and actually delivered on the AI promise - and that only worked on Nvidia cards.

DLSS 2.0 and RT actually don't interact super well with each other, but Nvidia very smartly figured out how to make them seem like they did. Instead of selling DLSS 2.0 as a way to make resolutions higher, it sold it as a way to make frame rates higher - as a tech that could recover the performance "lost" by enabling RT. So even though the two technologies actually fight each other a little bit under the hood, Nvidia managed to find a way to make them seem tied at the hip.

So Nvidia had features, but arguably, AMD had power and cost. But it was about to get worse for AMD. Even with the power of advanced GPUs, developers were having trouble pushing all those damn pixels with 4k everything, and so they started to use temporal upscaling - the same class of tech as DLSS 2 - everywhere. AMD released a best-in-class upscaler (FSR 2) which delivered similar results to DLSS without the tensor cores.

On paper that's great - it is great! - but it meant that all that extra power AMD had bet on was being used to replicate Nvidia features. Game X might run better on AMD than Nvidia out the gate, but then enable DLSS 2.0 and NVidia runs much better. AMD brings out FSR2 to match, but FSR2 itself eats the extra power that gave AMD the advantage in the first place. And it doesn't look quite as good without that AI to help.

Then came the RTX 40, and jaws dropped. Prices were awful because the advanced foundry nodes that GPUs had been rushing to for decades were getting more and more expensive, and without chiplets, Nvidia was carrying that cost on every single square millimeter on their new chip. This was exactly what AMD had expected, and why they invested in chiplet designs.

Months later and the RX 7000 series was revealed and prices... were just as bad. AMD had pulled it off, they had managed to build a chiplet GPU. But it turned out to be a very different problem than a chiplet CPU, and because of that, very little of the GPU could be built on a cheaper node, and thus, very little cost savings on this first version of the design.

And there were other problems with RX 7000 as well. AMD updated their cores to have some of Nvidia's advantages that made DLSS/RT fast - like dual issue compute, and accelerated matrix instructions. But the commitment to backwards compatibility was showing it's age, with lots of complexity in the front end making utilization of these features fall way short of their theoretical max.

But Nvidia does have chickens coming home to roost. AMD might not have nailed it, but they weren't wrong. Chiplets are the future, and Nvidia has to get there. Nvidia didn't skip the chiplet investment, they just delayed it. And in this time of surging AI products, AMD's chiplet design is paying off. They're able to put together custom data center products that combine several of their technologies extremely quickly.

When it comes to backwards compatibility, Nvidia and Nintendo have likely invested huge quantities of money to make it happen, and will probably have to do so again in 5 years. If their BC is emulation driven, then Nvidia is developing the software that make it possible for Nintendo to go to a different vendor in the future. AMD has gotten said BC nearly for free, and has locked in the other two console makers likely for a couple more generations.

AMD is also innovating, with the recent previews of Frame Gen technology that works in legacy games without patches. That's potentially a huge win for the PS6/Next Box, allowing 120fps modes for everything but also potentially a major win for those handheld PCs whose value proposition is often around being able to run last gen games in your hand.

And AMD is likely to dominant in the handheld PC space not just because they have top-tier PC CPU, but because again, this is a place where their chiplet tech has huge potential to pay off, with AMD able to deliver customized APUs extremely quickly, and with a low enough design cost that a customized APU is affordable even for products that don't sell millions of units.

It remains to be seen if AMD can deliver on the potential of chiplets, and can catch up on the feature space. But it also remains to be seen how much further Nvidia can take DLSS, with Frame Gen and Ray Reconstruction being the obvious evolutions of the tech. Nvidia has got a big roadblock with their move to chiplets, but AMD actually already has top-class machine learning hardware in their server offerings, and could catch up rapidly if they decide to go that path.

May you live in interesting times!

I think it's great how you take the time to explain something like this in such detail, it's genuinely interesting. Thank you.

And thank you to all of the other members who contribute to make this such a thread such an amazing read.

 

[Deleted member 887]

The real issue is that AMD spent a long time barely holding on financially.

Honestly, this is a better answer

 

[chocolate_supra]

I'm gonna actually try and answer this question, so buckle up

TL;DR: Nvidia bet on features, AMD bet on power. Features won in the market, which left AMD spending their extra power to simulate those features, leaving them with less power to go around, and features that aren't as good.

Every big bump in resolution roughly doubles how much detail a human eye can pick up, but roughly quadruples the number of pixels. And because it quadruples the pixels, it quadruples the amount of power it takes to put that stuff on screen. If you think about that for more than a minute, the problem becomes obvious - this shit can't go on forever.

And that resolution leap doesn't include making this pixels prettier. Not just advanced details but advanced effects, like higher quality lighting, reflections, etcetera. So you need to quadruple performance just to stand still. You need to do better than that to advance.

In every field except the GPU, those advances in performance have become extremely difficult. At some point, making CPU's faster got really hard, which is why they put more and more cores in every generation. GPUs happen to scale very well with adding more cores, so GPUs have dodged the wall that other system's have been hitting. But that won't last forever.

Both Nvidia and AMD clearly saw this writing on the wall. Neither of them (or Intel, in fact, but that's a tangent right now) misunderstood the problem. What happened next is that they tried two very different solutions.

AMD is a secondary player in the desktop space, with a lot of their core consumers being budget players. They absolutely dominate consoles, and have for the last 2 decades. They're a strong player in the data center, and they have a CPU product that dominates the industry, and is based on a technology called "chiplets" where they can mix and match parts from different foundries. That lets them rapidly customize products, while also manufacturing performance critical chunks of a chip with the most advanced but expensive tech, and less performance critical chunks on cheaper tech.

AMD's strategy was this - keep pursuing that classic gen-on-gen power, by iterating on their core design. Keep it backwards compatible for their console customers. Keep their data center and consumer segments different, but invest heavily in bringing their chiplet tech to GPUs. That will allow them to very quickly adapt products to the market without having to design new hardware from scratch each time, while also keeping costs down.

AMD saw the wall coming, laid down the gauntlet and said, fuck it, we're going to bust straight through that thing. It was smart and aggressive.

That... is not what Nvidia did. Nvidia decided that the only winning move was not to play. Nvidia decided that instead of pursuing more and more power they would pursue features. Nvidia added Ray Tracing, which doesn't make More Pixels, but does make Prettier Pixels. And because it's a relatively new tech for the consumer space, there is a much much longer road of innovation ahead of them, betting they can deliver huge leaps on the RT side while the traditional rasterization side slows down.

They didn't pursue chiplets, instead deciding to just make their datacenter designs and their consumer designs the same to reduce design costs. That meant putting AI hardware on consumer products, AI being another feature where huge leaps are still possible. And that meant finding a use for that AI hardware in the first place.
Which lead to AI assisted upscaling, and DLSS in the first place.

Early on, it looked like AMD had pulled it off. Low RT powered device couldn't deliver much, and few games took advantage of it without Nvidia throwing money at it. AMD figured out how to add basic RT to their hardware with minimal modification, instead of the huge investment Nvidia made. At "traditional" rendering, AMD was delivering better performance, and AI upscaling wasn't just bad, it invented new kinds of bad no one had ever seen before. Bad upscaling would miss detail, and just create blurrier images. DLSS 1.0 was instead finding detail that didn't exist and looked wrong adding bizarre details to images that made no sense. And it was expensive requiring a supercomputer to train an custom AI model for each game that wanted to use it.

Then 2020 happened. Control was out, and people started to see RT could really do, and because AMD had at least minimal support, RT modes started to become common in games, which of course ran better on Nvidia. And then came DLSS 2.0 which was a generic solution, easy to implement, that didn't have DLSS 1.0's problems and actually delivered on the AI promise - and that only worked on Nvidia cards.

DLSS 2.0 and RT actually don't interact super well with each other, but Nvidia very smartly figured out how to make them seem like they did. Instead of selling DLSS 2.0 as a way to make resolutions higher, it sold it as a way to make frame rates higher - as a tech that could recover the performance "lost" by enabling RT. So even though the two technologies actually fight each other a little bit under the hood, Nvidia managed to find a way to make them seem tied at the hip.

So Nvidia had features, but arguably, AMD had power and cost. But it was about to get worse for AMD. Even with the power of advanced GPUs, developers were having trouble pushing all those damn pixels with 4k everything, and so they started to use temporal upscaling - the same class of tech as DLSS 2 - everywhere. AMD released a best-in-class upscaler (FSR 2) which delivered similar results to DLSS without the tensor cores.

On paper that's great - it is great! - but it meant that all that extra power AMD had bet on was being used to replicate Nvidia features. Game X might run better on AMD than Nvidia out the gate, but then enable DLSS 2.0 and NVidia runs much better. AMD brings out FSR2 to match, but FSR2 itself eats the extra power that gave AMD the advantage in the first place. And it doesn't look quite as good without that AI to help.

Then came the RTX 40, and jaws dropped. Prices were awful because the advanced foundry nodes that GPUs had been rushing to for decades were getting more and more expensive, and without chiplets, Nvidia was carrying that cost on every single square millimeter on their new chip. This was exactly what AMD had expected, and why they invested in chiplet designs.

Months later and the RX 7000 series was revealed and prices... were just as bad. AMD had pulled it off, they had managed to build a chiplet GPU. But it turned out to be a very different problem than a chiplet CPU, and because of that, very little of the GPU could be built on a cheaper node, and thus, very little cost savings on this first version of the design.

And there were other problems with RX 7000 as well. AMD updated their cores to have some of Nvidia's advantages that made DLSS/RT fast - like dual issue compute, and accelerated matrix instructions. But the commitment to backwards compatibility was showing it's age, with lots of complexity in the front end making utilization of these features fall way short of their theoretical max.

But Nvidia does have chickens coming home to roost. AMD might not have nailed it, but they weren't wrong. Chiplets are the future, and Nvidia has to get there. Nvidia didn't skip the chiplet investment, they just delayed it. And in this time of surging AI products, AMD's chiplet design is paying off. They're able to put together custom data center products that combine several of their technologies extremely quickly.

When it comes to backwards compatibility, Nvidia and Nintendo have likely invested huge quantities of money to make it happen, and will probably have to do so again in 5 years. If their BC is emulation driven, then Nvidia is developing the software that make it possible for Nintendo to go to a different vendor in the future. AMD has gotten said BC nearly for free, and has locked in the other two console makers likely for a couple more generations.

AMD is also innovating, with the recent previews of Frame Gen technology that works in legacy games without patches. That's potentially a huge win for the PS6/Next Box, allowing 120fps modes for everything but also potentially a major win for those handheld PCs whose value proposition is often around being able to run last gen games in your hand.

And AMD is likely to dominant in the handheld PC space not just because they have top-tier PC CPU, but because again, this is a place where their chiplet tech has huge potential to pay off, with AMD able to deliver customized APUs extremely quickly, and with a low enough design cost that a customized APU is affordable even for products that don't sell millions of units.

It remains to be seen if AMD can deliver on the potential of chiplets, and can catch up on the feature space. But it also remains to be seen how much further Nvidia can take DLSS, with Frame Gen and Ray Reconstruction being the obvious evolutions of the tech. Nvidia has got a big roadblock with their move to chiplets, but AMD actually already has top-class machine learning hardware in their server offerings, and could catch up rapidly if they decide to go that path.

May you live in interesting times!

1. This was a very interesting read and made a ton of sense, even to a layman like me. Really appreciate you taking the time.

2. Even with this my brain found a way to relate it to cars and I think I need help yall

 

[Skittzo]

Prime 4 with raymond is gonna looks so nice

 

[Leviraps]

Most likely a stable 60, more in-game assets using objects and character models with higher polygon counts and greater texture resolution, greater draw distances and anisotropic filtering with none of that fizzy shadow cascading, possible global illumination with increased reflection ranges, then being able to run through Kakariko Village in the rain along with the Lost Woods, Faron jungle and Korok Forest without a noticeable frame dip

 

[snoota]

I'm gonna actually try and answer this question, so buckle up

TL;DR: Nvidia bet on features, AMD bet on power. Features won in the market, which left AMD spending their extra power to simulate those features, leaving them with less power to go around, and features that aren't as good.

Every big bump in resolution roughly doubles how much detail a human eye can pick up, but roughly quadruples the number of pixels. And because it quadruples the pixels, it quadruples the amount of power it takes to put that stuff on screen. If you think about that for more than a minute, the problem becomes obvious - this shit can't go on forever.

And that resolution leap doesn't include making this pixels prettier. Not just advanced details but advanced effects, like higher quality lighting, reflections, etcetera. So you need to quadruple performance just to stand still. You need to do better than that to advance.

In every field except the GPU, those advances in performance have become extremely difficult. At some point, making CPU's faster got really hard, which is why they put more and more cores in every generation. GPUs happen to scale very well with adding more cores, so GPUs have dodged the wall that other system's have been hitting. But that won't last forever.

Both Nvidia and AMD clearly saw this writing on the wall. Neither of them (or Intel, in fact, but that's a tangent right now) misunderstood the problem. What happened next is that they tried two very different solutions.

AMD is a secondary player in the desktop space, with a lot of their core consumers being budget players. They absolutely dominate consoles, and have for the last 2 decades. They're a strong player in the data center, and they have a CPU product that dominates the industry, and is based on a technology called "chiplets" where they can mix and match parts from different foundries. That lets them rapidly customize products, while also manufacturing performance critical chunks of a chip with the most advanced but expensive tech, and less performance critical chunks on cheaper tech.

AMD's strategy was this - keep pursuing that classic gen-on-gen power, by iterating on their core design. Keep it backwards compatible for their console customers. Keep their data center and consumer segments different, but invest heavily in bringing their chiplet tech to GPUs. That will allow them to very quickly adapt products to the market without having to design new hardware from scratch each time, while also keeping costs down.

AMD saw the wall coming, laid down the gauntlet and said, fuck it, we're going to bust straight through that thing. It was smart and aggressive.

That... is not what Nvidia did. Nvidia decided that the only winning move was not to play. Nvidia decided that instead of pursuing more and more power they would pursue features. Nvidia added Ray Tracing, which doesn't make More Pixels, but does make Prettier Pixels. And because it's a relatively new tech for the consumer space, there is a much much longer road of innovation ahead of them, betting they can deliver huge leaps on the RT side while the traditional rasterization side slows down.

They didn't pursue chiplets, instead deciding to just make their datacenter designs and their consumer designs the same to reduce design costs. That meant putting AI hardware on consumer products, AI being another feature where huge leaps are still possible. And that meant finding a use for that AI hardware in the first place.
Which lead to AI assisted upscaling, and DLSS in the first place.

Early on, it looked like AMD had pulled it off. Low RT powered device couldn't deliver much, and few games took advantage of it without Nvidia throwing money at it. AMD figured out how to add basic RT to their hardware with minimal modification, instead of the huge investment Nvidia made. At "traditional" rendering, AMD was delivering better performance, and AI upscaling wasn't just bad, it invented new kinds of bad no one had ever seen before. Bad upscaling would miss detail, and just create blurrier images. DLSS 1.0 was instead finding detail that didn't exist and looked wrong adding bizarre details to images that made no sense. And it was expensive requiring a supercomputer to train an custom AI model for each game that wanted to use it.

Then 2020 happened. Control was out, and people started to see RT could really do, and because AMD had at least minimal support, RT modes started to become common in games, which of course ran better on Nvidia. And then came DLSS 2.0 which was a generic solution, easy to implement, that didn't have DLSS 1.0's problems and actually delivered on the AI promise - and that only worked on Nvidia cards.

DLSS 2.0 and RT actually don't interact super well with each other, but Nvidia very smartly figured out how to make them seem like they did. Instead of selling DLSS 2.0 as a way to make resolutions higher, it sold it as a way to make frame rates higher - as a tech that could recover the performance "lost" by enabling RT. So even though the two technologies actually fight each other a little bit under the hood, Nvidia managed to find a way to make them seem tied at the hip.

So Nvidia had features, but arguably, AMD had power and cost. But it was about to get worse for AMD. Even with the power of advanced GPUs, developers were having trouble pushing all those damn pixels with 4k everything, and so they started to use temporal upscaling - the same class of tech as DLSS 2 - everywhere. AMD released a best-in-class upscaler (FSR 2) which delivered similar results to DLSS without the tensor cores.

On paper that's great - it is great! - but it meant that all that extra power AMD had bet on was being used to replicate Nvidia features. Game X might run better on AMD than Nvidia out the gate, but then enable DLSS 2.0 and NVidia runs much better. AMD brings out FSR2 to match, but FSR2 itself eats the extra power that gave AMD the advantage in the first place. And it doesn't look quite as good without that AI to help.

Then came the RTX 40, and jaws dropped. Prices were awful because the advanced foundry nodes that GPUs had been rushing to for decades were getting more and more expensive, and without chiplets, Nvidia was carrying that cost on every single square millimeter on their new chip. This was exactly what AMD had expected, and why they invested in chiplet designs.

Months later and the RX 7000 series was revealed and prices... were just as bad. AMD had pulled it off, they had managed to build a chiplet GPU. But it turned out to be a very different problem than a chiplet CPU, and because of that, very little of the GPU could be built on a cheaper node, and thus, very little cost savings on this first version of the design.

And there were other problems with RX 7000 as well. AMD updated their cores to have some of Nvidia's advantages that made DLSS/RT fast - like dual issue compute, and accelerated matrix instructions. But the commitment to backwards compatibility was showing it's age, with lots of complexity in the front end making utilization of these features fall way short of their theoretical max.

But Nvidia does have chickens coming home to roost. AMD might not have nailed it, but they weren't wrong. Chiplets are the future, and Nvidia has to get there. Nvidia didn't skip the chiplet investment, they just delayed it. And in this time of surging AI products, AMD's chiplet design is paying off. They're able to put together custom data center products that combine several of their technologies extremely quickly.

When it comes to backwards compatibility, Nvidia and Nintendo have likely invested huge quantities of money to make it happen, and will probably have to do so again in 5 years. If their BC is emulation driven, then Nvidia is developing the software that make it possible for Nintendo to go to a different vendor in the future. AMD has gotten said BC nearly for free, and has locked in the other two console makers likely for a couple more generations.

AMD is also innovating, with the recent previews of Frame Gen technology that works in legacy games without patches. That's potentially a huge win for the PS6/Next Box, allowing 120fps modes for everything but also potentially a major win for those handheld PCs whose value proposition is often around being able to run last gen games in your hand.

And AMD is likely to dominant in the handheld PC space not just because they have top-tier PC CPU, but because again, this is a place where their chiplet tech has huge potential to pay off, with AMD able to deliver customized APUs extremely quickly, and with a low enough design cost that a customized APU is affordable even for products that don't sell millions of units.

It remains to be seen if AMD can deliver on the potential of chiplets, and can catch up on the feature space. But it also remains to be seen how much further Nvidia can take DLSS, with Frame Gen and Ray Reconstruction being the obvious evolutions of the tech. Nvidia has got a big roadblock with their move to chiplets, but AMD actually already has top-class machine learning hardware in their server offerings, and could catch up rapidly if they decide to go that path.

May you live in interesting times!

RX 7000 really suffered from AMD failing to meet clock & perf/watt expectations. It's very evidently clear that in BoM costs that each SKU was intended to cost comparably to the lower tier SKU from Nvidia whilst competing in Raster/power draw to its higher tier equivalent (So N31 costing comparable to AD103, N32 costing similar to AD104, & N33 competing against AD107 in pricing), but it instead ended up with an absolute trash v/f power curve & VOPD not being as consistently useful for gaming performance as Nvidia's split FP32 + INT32/FP32 pipeline. It's all around made worse by Lovelace being downright superior from a physdes design standpoint, albeit having the benefit of a massive nodeshrink a la Maxwell -> Pascal.

 

[Giancarlo]

Nintendo underpays even for their unknown VOs. The video game industry as a whole aggressively pursues non-Union talent, but Nintendo especially is known for pursuing amateur talent. Charles is a little unusual for being a professional actor, probably because he didn't start working on Mario for video games at all, but for their trade shows (which are non-union gigs) and only got kept on for the video games because Miyamoto liked him personally.

This seems to be changing a bit, with "proper" voice actors doing more and more work. But I can't imagine them going with Pratt, not because of the cost, but the scheduling. No way would Pratt's agent let him sell perpetuity rights like Martinet did, meaning voice clips couldn't be used in future games, so they'd have to find time to schedule voice recording sessions around a Hollywood actor's schedule, on a game development timeline. Nightmare

Nintendo focus in on gameplay, not huge cinematic games like Sony focus is, they would rather do a fully focused franchises/games, rater then adding plots

 

[eye0h]

You guys do realize the Mario movie is one of the most successful animated movies of all time, right?

never heard of it

 

[chocolate_supra]

Prime 4 with raymond is gonna looks so nice

 

[ShadowFox08]

Just to be clear, I'm not saying that Nvidia's the only company making suggestions to Nintendo.

My point is that just because Nvidia mentioned that Jetson Orin NX has 16 GB of RAM doesn't mean Nintendo's going to automatically do likewise for retail hardware. As I've said before, Nintendo makes the final decision, not Nvidia, nor any company making suggestions to Nintendo.

---

Speaking of RAM, I don't think the RAM amount's going to be a bottleneck for Nintendo's new hardware. I think like with the Nintendo Switch, the CPU frequency and the RAM bandwidth are going to be the biggest bottlenecks with Nintendo's new hardware, although I don't believe to the same degree as with the Nintendo Switch.

Agreed. We'll have more RAM available for games than last gen PS4 Pro/One X, and current gen Series S even.

 

[bossmeran]

I'm gonna actually try and answer this question, so buckle up

TL;DR: Nvidia bet on features, AMD bet on power. Features won in the market, which left AMD spending their extra power to simulate those features, leaving them with less power to go around, and features that aren't as good.

Every big bump in resolution roughly doubles how much detail a human eye can pick up, but roughly quadruples the number of pixels. And because it quadruples the pixels, it quadruples the amount of power it takes to put that stuff on screen. If you think about that for more than a minute, the problem becomes obvious - this shit can't go on forever.

And that resolution leap doesn't include making this pixels prettier. Not just advanced details but advanced effects, like higher quality lighting, reflections, etcetera. So you need to quadruple performance just to stand still. You need to do better than that to advance.

In every field except the GPU, those advances in performance have become extremely difficult. At some point, making CPU's faster got really hard, which is why they put more and more cores in every generation. GPUs happen to scale very well with adding more cores, so GPUs have dodged the wall that other system's have been hitting. But that won't last forever.

Both Nvidia and AMD clearly saw this writing on the wall. Neither of them (or Intel, in fact, but that's a tangent right now) misunderstood the problem. What happened next is that they tried two very different solutions.

AMD is a secondary player in the desktop space, with a lot of their core consumers being budget players. They absolutely dominate consoles, and have for the last 2 decades. They're a strong player in the data center, and they have a CPU product that dominates the industry, and is based on a technology called "chiplets" where they can mix and match parts from different foundries. That lets them rapidly customize products, while also manufacturing performance critical chunks of a chip with the most advanced but expensive tech, and less performance critical chunks on cheaper tech.

AMD's strategy was this - keep pursuing that classic gen-on-gen power, by iterating on their core design. Keep it backwards compatible for their console customers. Keep their data center and consumer segments different, but invest heavily in bringing their chiplet tech to GPUs. That will allow them to very quickly adapt products to the market without having to design new hardware from scratch each time, while also keeping costs down.

AMD saw the wall coming, laid down the gauntlet and said, fuck it, we're going to bust straight through that thing. It was smart and aggressive.

That... is not what Nvidia did. Nvidia decided that the only winning move was not to play. Nvidia decided that instead of pursuing more and more power they would pursue features. Nvidia added Ray Tracing, which doesn't make More Pixels, but does make Prettier Pixels. And because it's a relatively new tech for the consumer space, there is a much much longer road of innovation ahead of them, betting they can deliver huge leaps on the RT side while the traditional rasterization side slows down.

They didn't pursue chiplets, instead deciding to just make their datacenter designs and their consumer designs the same to reduce design costs. That meant putting AI hardware on consumer products, AI being another feature where huge leaps are still possible. And that meant finding a use for that AI hardware in the first place.
Which lead to AI assisted upscaling, and DLSS in the first place.

Early on, it looked like AMD had pulled it off. Low RT powered device couldn't deliver much, and few games took advantage of it without Nvidia throwing money at it. AMD figured out how to add basic RT to their hardware with minimal modification, instead of the huge investment Nvidia made. At "traditional" rendering, AMD was delivering better performance, and AI upscaling wasn't just bad, it invented new kinds of bad no one had ever seen before. Bad upscaling would miss detail, and just create blurrier images. DLSS 1.0 was instead finding detail that didn't exist and looked wrong adding bizarre details to images that made no sense. And it was expensive requiring a supercomputer to train an custom AI model for each game that wanted to use it.

Then 2020 happened. Control was out, and people started to see RT could really do, and because AMD had at least minimal support, RT modes started to become common in games, which of course ran better on Nvidia. And then came DLSS 2.0 which was a generic solution, easy to implement, that didn't have DLSS 1.0's problems and actually delivered on the AI promise - and that only worked on Nvidia cards.

DLSS 2.0 and RT actually don't interact super well with each other, but Nvidia very smartly figured out how to make them seem like they did. Instead of selling DLSS 2.0 as a way to make resolutions higher, it sold it as a way to make frame rates higher - as a tech that could recover the performance "lost" by enabling RT. So even though the two technologies actually fight each other a little bit under the hood, Nvidia managed to find a way to make them seem tied at the hip.

So Nvidia had features, but arguably, AMD had power and cost. But it was about to get worse for AMD. Even with the power of advanced GPUs, developers were having trouble pushing all those damn pixels with 4k everything, and so they started to use temporal upscaling - the same class of tech as DLSS 2 - everywhere. AMD released a best-in-class upscaler (FSR 2) which delivered similar results to DLSS without the tensor cores.

On paper that's great - it is great! - but it meant that all that extra power AMD had bet on was being used to replicate Nvidia features. Game X might run better on AMD than Nvidia out the gate, but then enable DLSS 2.0 and NVidia runs much better. AMD brings out FSR2 to match, but FSR2 itself eats the extra power that gave AMD the advantage in the first place. And it doesn't look quite as good without that AI to help.

Then came the RTX 40, and jaws dropped. Prices were awful because the advanced foundry nodes that GPUs had been rushing to for decades were getting more and more expensive, and without chiplets, Nvidia was carrying that cost on every single square millimeter on their new chip. This was exactly what AMD had expected, and why they invested in chiplet designs.

Months later and the RX 7000 series was revealed and prices... were just as bad. AMD had pulled it off, they had managed to build a chiplet GPU. But it turned out to be a very different problem than a chiplet CPU, and because of that, very little of the GPU could be built on a cheaper node, and thus, very little cost savings on this first version of the design.

And there were other problems with RX 7000 as well. AMD updated their cores to have some of Nvidia's advantages that made DLSS/RT fast - like dual issue compute, and accelerated matrix instructions. But the commitment to backwards compatibility was showing it's age, with lots of complexity in the front end making utilization of these features fall way short of their theoretical max.

But Nvidia does have chickens coming home to roost. AMD might not have nailed it, but they weren't wrong. Chiplets are the future, and Nvidia has to get there. Nvidia didn't skip the chiplet investment, they just delayed it. And in this time of surging AI products, AMD's chiplet design is paying off. They're able to put together custom data center products that combine several of their technologies extremely quickly.

When it comes to backwards compatibility, Nvidia and Nintendo have likely invested huge quantities of money to make it happen, and will probably have to do so again in 5 years. If their BC is emulation driven, then Nvidia is developing the software that make it possible for Nintendo to go to a different vendor in the future. AMD has gotten said BC nearly for free, and has locked in the other two console makers likely for a couple more generations.

AMD is also innovating, with the recent previews of Frame Gen technology that works in legacy games without patches. That's potentially a huge win for the PS6/Next Box, allowing 120fps modes for everything but also potentially a major win for those handheld PCs whose value proposition is often around being able to run last gen games in your hand.

And AMD is likely to dominant in the handheld PC space not just because they have top-tier PC CPU, but because again, this is a place where their chiplet tech has huge potential to pay off, with AMD able to deliver customized APUs extremely quickly, and with a low enough design cost that a customized APU is affordable even for products that don't sell millions of units.

It remains to be seen if AMD can deliver on the potential of chiplets, and can catch up on the feature space. But it also remains to be seen how much further Nvidia can take DLSS, with Frame Gen and Ray Reconstruction being the obvious evolutions of the tech. Nvidia has got a big roadblock with their move to chiplets, but AMD actually already has top-class machine learning hardware in their server offerings, and could catch up rapidly if they decide to go that path.

May you live in interesting times!

Wow just wow. I love this stuff it's fascinating. So it's all a matter of luck via business decisions and prioritising different paths to expanding business. You should genuinely start a YouTube channel.

Maybe chiplets will prove to give AMD an advantage in the gaming space but ATM Nvidia is the place to be. AMD apparently said that RDNA4 has "advancements in AI and ML" so maybe that'll help them catch up. Probably just marketing though. Who knows where intel is lmao. I really do hope AMD catches up, especially in the handheld PC space since I want all portables to be as powerful as possible so us portable fans can have ideal hardware whether PC or console. I vastly prefer playing handhelds!

Do you know when AMD and Nvidia tend to announce and show off their developments and specifically SoCs? Wanna keep an eye on it to see how console and PC handhelds develop. I think the next few years will be huge for handhelds as we're getting to a point where they're within spitting distance of home consoles to the layman.

 

[Xbro]

I'm gonna actually try and answer this question, so buckle up

TL;DR: Nvidia bet on features, AMD bet on power. Features won in the market, which left AMD spending their extra power to simulate those features, leaving them with less power to go around, and features that aren't as good.

Every big bump in resolution roughly doubles how much detail a human eye can pick up, but roughly quadruples the number of pixels. And because it quadruples the pixels, it quadruples the amount of power it takes to put that stuff on screen. If you think about that for more than a minute, the problem becomes obvious - this shit can't go on forever.

And that resolution leap doesn't include making this pixels prettier. Not just advanced details but advanced effects, like higher quality lighting, reflections, etcetera. So you need to quadruple performance just to stand still. You need to do better than that to advance.

In every field except the GPU, those advances in performance have become extremely difficult. At some point, making CPU's faster got really hard, which is why they put more and more cores in every generation. GPUs happen to scale very well with adding more cores, so GPUs have dodged the wall that other system's have been hitting. But that won't last forever.

Both Nvidia and AMD clearly saw this writing on the wall. Neither of them (or Intel, in fact, but that's a tangent right now) misunderstood the problem. What happened next is that they tried two very different solutions.

AMD is a secondary player in the desktop space, with a lot of their core consumers being budget players. They absolutely dominate consoles, and have for the last 2 decades. They're a strong player in the data center, and they have a CPU product that dominates the industry, and is based on a technology called "chiplets" where they can mix and match parts from different foundries. That lets them rapidly customize products, while also manufacturing performance critical chunks of a chip with the most advanced but expensive tech, and less performance critical chunks on cheaper tech.

AMD's strategy was this - keep pursuing that classic gen-on-gen power, by iterating on their core design. Keep it backwards compatible for their console customers. Keep their data center and consumer segments different, but invest heavily in bringing their chiplet tech to GPUs. That will allow them to very quickly adapt products to the market without having to design new hardware from scratch each time, while also keeping costs down.

AMD saw the wall coming, laid down the gauntlet and said, fuck it, we're going to bust straight through that thing. It was smart and aggressive.

That... is not what Nvidia did. Nvidia decided that the only winning move was not to play. Nvidia decided that instead of pursuing more and more power they would pursue features. Nvidia added Ray Tracing, which doesn't make More Pixels, but does make Prettier Pixels. And because it's a relatively new tech for the consumer space, there is a much much longer road of innovation ahead of them, betting they can deliver huge leaps on the RT side while the traditional rasterization side slows down.

They didn't pursue chiplets, instead deciding to just make their datacenter designs and their consumer designs the same to reduce design costs. That meant putting AI hardware on consumer products, AI being another feature where huge leaps are still possible. And that meant finding a use for that AI hardware in the first place.
Which lead to AI assisted upscaling, and DLSS in the first place.

Early on, it looked like AMD had pulled it off. Low RT powered device couldn't deliver much, and few games took advantage of it without Nvidia throwing money at it. AMD figured out how to add basic RT to their hardware with minimal modification, instead of the huge investment Nvidia made. At "traditional" rendering, AMD was delivering better performance, and AI upscaling wasn't just bad, it invented new kinds of bad no one had ever seen before. Bad upscaling would miss detail, and just create blurrier images. DLSS 1.0 was instead finding detail that didn't exist and looked wrong adding bizarre details to images that made no sense. And it was expensive requiring a supercomputer to train an custom AI model for each game that wanted to use it.

Then 2020 happened. Control was out, and people started to see RT could really do, and because AMD had at least minimal support, RT modes started to become common in games, which of course ran better on Nvidia. And then came DLSS 2.0 which was a generic solution, easy to implement, that didn't have DLSS 1.0's problems and actually delivered on the AI promise - and that only worked on Nvidia cards.

DLSS 2.0 and RT actually don't interact super well with each other, but Nvidia very smartly figured out how to make them seem like they did. Instead of selling DLSS 2.0 as a way to make resolutions higher, it sold it as a way to make frame rates higher - as a tech that could recover the performance "lost" by enabling RT. So even though the two technologies actually fight each other a little bit under the hood, Nvidia managed to find a way to make them seem tied at the hip.

So Nvidia had features, but arguably, AMD had power and cost. But it was about to get worse for AMD. Even with the power of advanced GPUs, developers were having trouble pushing all those damn pixels with 4k everything, and so they started to use temporal upscaling - the same class of tech as DLSS 2 - everywhere. AMD released a best-in-class upscaler (FSR 2) which delivered similar results to DLSS without the tensor cores.

On paper that's great - it is great! - but it meant that all that extra power AMD had bet on was being used to replicate Nvidia features. Game X might run better on AMD than Nvidia out the gate, but then enable DLSS 2.0 and NVidia runs much better. AMD brings out FSR2 to match, but FSR2 itself eats the extra power that gave AMD the advantage in the first place. And it doesn't look quite as good without that AI to help.

Then came the RTX 40, and jaws dropped. Prices were awful because the advanced foundry nodes that GPUs had been rushing to for decades were getting more and more expensive, and without chiplets, Nvidia was carrying that cost on every single square millimeter on their new chip. This was exactly what AMD had expected, and why they invested in chiplet designs.

Months later and the RX 7000 series was revealed and prices... were just as bad. AMD had pulled it off, they had managed to build a chiplet GPU. But it turned out to be a very different problem than a chiplet CPU, and because of that, very little of the GPU could be built on a cheaper node, and thus, very little cost savings on this first version of the design.

And there were other problems with RX 7000 as well. AMD updated their cores to have some of Nvidia's advantages that made DLSS/RT fast - like dual issue compute, and accelerated matrix instructions. But the commitment to backwards compatibility was showing it's age, with lots of complexity in the front end making utilization of these features fall way short of their theoretical max.

But Nvidia does have chickens coming home to roost. AMD might not have nailed it, but they weren't wrong. Chiplets are the future, and Nvidia has to get there. Nvidia didn't skip the chiplet investment, they just delayed it. And in this time of surging AI products, AMD's chiplet design is paying off. They're able to put together custom data center products that combine several of their technologies extremely quickly.

When it comes to backwards compatibility, Nvidia and Nintendo have likely invested huge quantities of money to make it happen, and will probably have to do so again in 5 years. If their BC is emulation driven, then Nvidia is developing the software that make it possible for Nintendo to go to a different vendor in the future. AMD has gotten said BC nearly for free, and has locked in the other two console makers likely for a couple more generations.

AMD is also innovating, with the recent previews of Frame Gen technology that works in legacy games without patches. That's potentially a huge win for the PS6/Next Box, allowing 120fps modes for everything but also potentially a major win for those handheld PCs whose value proposition is often around being able to run last gen games in your hand.

And AMD is likely to dominant in the handheld PC space not just because they have top-tier PC CPU, but because again, this is a place where their chiplet tech has huge potential to pay off, with AMD able to deliver customized APUs extremely quickly, and with a low enough design cost that a customized APU is affordable even for products that don't sell millions of units.

It remains to be seen if AMD can deliver on the potential of chiplets, and can catch up on the feature space. But it also remains to be seen how much further Nvidia can take DLSS, with Frame Gen and Ray Reconstruction being the obvious evolutions of the tech. Nvidia has got a big roadblock with their move to chiplets, but AMD actually already has top-class machine learning hardware in their server offerings, and could catch up rapidly if they decide to go that path.

May you live in interesting times!

How long do you suppose it would take for Nvidia to implement their chiplet solution? Seems like they've been moving pretty rapidly the past few years.

 

[GPKs3]

Everybody Loves Metroid

 

[Skittzo]

Everybody Loves Sylux

 

[Goodtwin]

For me, yeah. There aren't a lot of details, but what's leaked out matches what I thought would be possible with that combination of performance and Nvidia hardware. So I feel like the high-level estimates are true

I would assume you are expecting Drake to be on 4N at this point? I think that is the only way the high end estimates are possible.

 

[Concernt]

I would assume you are expecting Drake to be on 4N at this point? I think that is the only way the high end estimates are possible.

That's what pretty much everyone is assuming right now. Not much would make sense if it wasn't.

 

[SoldierDelta]

FF7IIR on Switch 2 looking like Intergrade is plausible as the differences include resolution, lighting and texture improvements.

From our understanding of Switch 2 as a PS4+, it could render a 1080p image with bells and whistles and AI upscale to 1440p/2160p, making a sharper image comparable to the PS5 version.

So this specific claim by Hero isn't out of line with what we anticipate.

Ah, thank you.

With the understanding that basically every game will be using DLSS, FF7RI looking as good as the PS5 version isn't an unreasonable claim and could be an early title depending on Square's contract with Sony around the FF7 Remakes. Granted Hero's leak is still disputed among SE leakers so we'll have to wait and see.

 

[SiG]

540p to 1080p DLSS handheld
720 to 1440p DLSS Docked

More like 720p to 4k via Ultra Performance. It makes no sense to upscale to 1440p as that isn't a standard TV resolution.

 

[bossmeran]

The real issue is that AMD spent a long time barely holding on financially. The Bulldozer CPU architecture (and its successors) were disastrous for the company, and they lost almost all of their market share in the (high margin) server sector, and could only hang on to a small part the (low margin) entry level of the PC space. The GPU business wasn't doing completely terribly at the time, but CPU sales have always been the core of AMD's revenue stream, so they were struggling to make a profit as a company.

What do you do when you're struggling to make a profit and there's no quick way to increase sales? Cut back on any expenses that don't have a direct path to profitability. That means dropping exploratory R&D on long-term technologies like hardware ray tracing and AI. It also meant cancelling their planned "K12" ARM CPU core to focus on Zen instead (a sensible move in retrospect). AMD spent some very lean years working on a slim R&D budget that could only really justify straight-forward technological advances, meaning new CPU and GPU architectures that do the same thing, but faster. They couldn't justify the spend on something like hardware ray tracing that's not guaranteed to pay off.

Even their biggest innovation of the past few years was largely motivated by minimising R&D cost and risk. The reason they started using chiplets when Intel was still entirely focused on monolithic chips is that designing and taping out chips on leading-edge nodes costs a lot of money, money which AMD didn't have at the time. By moving to a chiplet approach, AMD could tape out just one chip on a leading-edge node, plus one I/O die on an older, cheaper node, and cover everything from entry-level desktops to 64-core servers. They were looking to do with one die what Intel were doing with 5 or 6, which they did quite successfully.

It was only really in 2019 that AMD's financials started to turn around. In 2019, AMD's revenue was $6.73 billion, by 2022 that had gone up to $23.6 billion. The payoff of the Zen architecture was slow, because although home PC builders adopted Ryzen pretty quickly, it took time to convince OEMs/server/HPC customers/etc. to actually consider AMD chips again. They're now in a much better place financially, so they can start investing in more of that fundamental R&D again, but it takes a long time to start up that kind of research. You have to hire experts in the area, you have to work out all the low-level fundamental details before you can start designing hardware, and even when you've designed the hardware it will be a year or two before it's in anyone's hands.

In terms of AI acceleration, I'd say AMD is actually in a pretty good place from a hardware point of view. They've been shipping HPC chips with matrix cores (basically the same thing as tensor cores) for several generations now, although initially they were more focussed on higher-precision work for HPC applications (ie really good FP64 performance). For pure AI use-cases, they seem to have made significant improvements, and it's quite possible that AMD's new MI300X flat-out outperforms Nvidia's H100 with equivalently optimised software.

Therein lies the problem, though, software. Nvidia spent a lot of time and money not just developing AI hardware, but developing the software packages and making sure they're widely used. Almost any AI software stack you'll find out there has far better support for Nvidia hardware than AMD (if it runs on AMD hardware at all), and that's a big hill AMD have to climb. One thing that's going for them is that, with the spike in demand for AI hardware, the industry is suddenly becoming a lot more interested in making sure they have other hardware options than just Nvidia. Nvidia is charging through the nose for H100s, which is making them a lot of money, but creates a big incentive for AI software that works well across a range of hardware, which AMD will look to capitalise on.

On the consumer side, AMD have added matrix cores to their GPUs starting with the RDNA3 architecture, but as yet haven't been using them in games. They didn't really publish performance figures on them, but this blog post states they can get 512 ops/clock/CU for FP16 and BF16. That would come to about 122 Tflops on the RX 7900XTX, which is about the same as an RTX 4070, which is to say more than enough for something like DLSS. Again, there's a lack of software there on AMD's side, which comes from them not having the head-start that Nvidia had. I'd wager we'll see a AI-based version of FSR at some point in the next few years, though.

On the ray tracing side of things, it's a trickier problem to solve. They added hardware triangle intersection testing pretty quickly, but it's a relatively simple bit of circuitry, so it was an easy win for them. I have no doubt they're working on hardware accelerated BVH traversal, but it's not as easy a problem to solve in hardware as triangle intersection testing, and I'm betting Nvidia were working on it for a long time before actually launching Turing. I'd expect to see it in the next generation or two of AMD GPUs, at which point it will probably close the gap significantly.

TLDR: AMD had no money for ages so couldn't afford R&D on things like RT and AI. They now have money, so can fund this R&D, but it takes time.

Awesome write up. So seems they will probably close the gap at some point. I hope they catch up since the competition will give us all better hardware!

 

[ILikeFeet]

How long do you suppose it would take for Nvidia to implement their chiplet solution? Seems like they've been moving pretty rapidly the past few years.

supposed Nvidia can do it now. they have research documents on it from years ago and publicly commented that it's a developed tool in their toolbox. I don't think they'll move to chiplets until they absolutely maximized monoliths first. there are problems with substrates capacity that chiplets are suffering from. all the chiplets are put onto a substrate and there's just not enough to go around. that's why their AI products are on backorder and probably why AMD can't ramp up volume

Ah, thank you.

With the understanding that basically every game will be using DLSS, FF7RI looking as good as the PS5 version isn't an unreasonable claim and could be an early title depending on Square's contract with Sony around the FF7 Remakes. Granted Hero's leak is still disputed among SE leakers so we'll have to wait and see.

SE's contract with Sony should be over by now for FF7R1

 

[ArchedThunder]

More like 720p to 4k via Ultra Performance. It makes no sense to upscale to 1440p as that isn't a standard TV resolution.

Upscaling to 1440p wouldn’t mean it’d push a 1440p signal, it’d still send a signal to the TV matching whatever your set output resolution is on the system. It’s the same as all the sub 4k games on other systems.

 

[RochHoch]

You guys do realize the Mario movie is one of the most successful animated movies of all time, right?

Who cares

Nintendo is the last company that would let a successful film adaptation dictate their craft

 

[SiG]

Upscaling to 1440p wouldn’t mean it’d push a 1440p signal, it’d still send a signal to the TV matching whatever your set output resolution is on the system.

I meant using DLSS to upscale to 1440p doesn't make any sense.

The ending output can be 4k, and adjust its upscale quality reults on the fly depending on load (Ultra Performance <-> Performance <-> Balanced).
720p to 4k at 60fps is definitely doable.

 

[Goodtwin]

It would be cool if the quality settings were directly tied to the users output resolution. For example, if the user is gaming on a 1080p TV instead of 4K and the game renders internally at 720p, they would get DLSS Quality mode instead of Performance mode.

 

[ArchedThunder]

I meant using DLSS to upscale to 1440p doesn't make any sense.

The ending output can be 4k, and adjust its upscale quality reults on the fly depending on load (Ultra Performance <-> Performance <-> Balanced).
720p to 4k at 60fps is definitely doable.

1440p is still cheaper than 4K, and devs are always going to be bumping up against the system’s max capabilities.

 

[Deleted member 887]

. Who knows where intel is lmao.

Intel is actually the most interesting for me.

Intel has bet on a ground up DirectX 12 design with no legacy hardware. Their bet is that they can deliver enough performance at lower cost, that they can get legacy games working with software alone.

Right now that puts them in a weird place. Their legacy game support is very bad. On modern games, they’re probably the best budget card, period - but since they only support modern CPU and motherboards, they’re not a good choice for a budget gaming PC build.

Their DLSS/FSR equivalent is AI accelerated like Nvidia, but it’s optional so it runs everywhere like FSR. And it generally looks better.

Intel has a strong CPU, unlike Nvidia, and have been delivering “integrated” graphics for a while. They’re not there yet, but I think they’re laying the ground work for coming for AMD

 

[Gerald]

2GHz would be the best for CPU, btw after RGT video,Is there a chance that Nvidia upgrade the T239 Drake architecture to Ada or would it have to be a completely new chip?

 

[ILikeFeet]

It would be cool if the quality settings were directly tied to the users output resolution. For example, if the user is gaming on a 1080p TV instead of 4K and the game renders internally at 720p, they would get DLSS Quality mode instead of Performance mode.

Immortals of Aveum does this. it...doesn't really help. the game's performance is kinda weird

2GHz would be the best for CPU, btw after RTG video,Is there a chance that Nvidia upgrade the T239 Drake architecture to Ada or would it have to be a completely new chip?

no. that video just shows RTG is still as useless as he's been in the past.

changing anything like the cpu or gpu arches would have to be a completely new chip and would get a new name

 

[BDGAME]

Talking about the demos, first about Zelda. There are people that think it's garantee to be a retro compatible console because of that. But Nintendo already utilize Zelda in a not playable demo before:



Demos of games that exist, with improved images, but that don't represent the final product, are nothing new either:



So, that Zelda is, like the reports says, probably only a Tech demo, not necessary proof of anything.

The Matrix Demo, on other hand, that is where we can get your hopes on. Last generation we had that demo that demonstrate the power of UE4:

That demo was so important to show the power of that gen hardwares that Nvidia choose it to demonstrate the power of Tegra X1. Plus, we can saw the differences of Tegra for a PS4 on it:



That demo showed to us what that gen can achieve and what the Nintendo Switch can do. You can see a lot of losses on Tegra demo, but the core is all there.

The Matrix demo can be the same thing. Probably it will have some losses from Ps5's version. Maybe less cars, draw distance or particles. Maybe more load times, but, if the end, it still the same demo and can run well, then we can hope for current gen ports can become something more common on the next gen switch.

 

[KeijiKG]

Not that I care about any of this, but X is far from representative of "the general populace".

You cared enough to respond and really that doesn't change the fact that the internet has other places where that split exists. So it's either you don't care about the sample set or you think not enough of it matters.

 

[Foltzie]

Do chiplets make sense for mobile and mobile sized devices? Do you sacrifice some performance by using an interposer to tie the chipsets together?

 

[KeijiKG]

You guys do realize the Mario movie is one of the most successful animated movies of all time, right?

Didn't take Frozen II's crown and got slaughtered by Barbie.

It's Mariover.