That's probably where thr sdk comes in. Getting your game to run on the new api in a simulated environment before you get actual hardware would make the process smoother. The 4k thing was probably based on being told that the need device will be capable of 4k output and projected performance
-
Hey everyone, staff have documented a list of banned content and subject matter that we feel are not consistent with site values, and don't make sense to host discussion of on Famiboards. This list (and the relevant reasoning per item) is viewable here.
-
Do you have audio editing experience and want to help out with the Famiboards Discussion Club Podcast? If so, we're looking for help and would love to have you on the team! Just let us know in the Podcast Thread if you are interested!
StarTopic Future Nintendo Hardware & Technology Speculation & Discussion |ST| (Read the staff posts before commenting!)
- Thread starter Dakhil
- Start date
Kise Ryota
Moblin
Has anything leaked about these two recently? Are 16GB of RAM confirmed yet? Do we already have solid information on what to expect in terms of storage? I haven't read this thread much in the last month.
neither of these things are known. all we know about ram is that it has to fit a 128-bit bus on the SoC. so the options range from 8GB to 32GB
Gotdatmoneyy
Shriekbat
Nothing ram wise is confirmed outside of a 128 bit bus.
The odds this thing has 16GB of RAM are low though. Gonna be 8GB or perhaps 12GB.
This is meaningless, for the most part. Nintendo forecast is almost always to lope bc it is super conservative, especially at the start of the year. Me and other investors assume much higher numbers for software and profit than company guidance.
They will even admit that in calls.
Last edited:
D
Deleted member 887
Guest
The Forbes 2000 annually ranks companies by size, based on their real assets, their sales that year, and their market value.
The Switch was a nice bump in power over the Wii U. It was not what we would, in console space, traditionally call a generational leap in performance. It could not readily achieve a 2x bump in resolution (720p to 1080p). A 2x bump represents the gap between the PS4 and the PS4 Pro before checkerboard rendering ups them to 4k.
Nintendo has known since day 1 how powerful their hardware is. They've known better than anyone else, and I'm sure had the most clear-headed impressions of how that hardware was going to age. They were well aware that they were not making a Wii or a Wii U where they could say to themselves that their control scheme would open up gameplay possibilities that would insulate them from industry movement.
Meanwhile, Pokemon or Sonic seem, to me, to be the absolute worst indicators that new hardware is needed.
Arceus and Scarlet and Violet are games that don't show a game company straining against limited hardware, they show a game company straining to take advantage of much more powerful hardware than they are used to, willing to ship embarrassingly poor performers. Those things won't change with more powerful hardware.
Sonic is a deeply cross-gen game. This is not a cut down PS5 game, this is a PS4 game that is nicely enhanced on the current gen consoles in a way that many other cross-gen games are. Sonic is not an indication that the industry has moved on and left the Switch behind, it is the exact opposite - the industry is continuing to target the consoles that were released before the Switch with good experiences, and what would have been a "miracle port" in the early days of the Switch is now a day and date release.
To sum up - Nintendo isn't resting on it's laurels thinking its hardware is timeless. They've known there was a ticking clock since the beginning, and are likely surprised by how well their hardware is holding up.
I expect you're right about 1080p, but I doubt it's because they found 25% extra performance optimizing the engine. I suspect it's temporal upscaling.
Game wise, haven’t we had ratings leaks for Biomutant, Batman Arkham and Borderlands 3 yet we’ve seen nothing of them. Could they be in limbo waiting for the next Nintendo console which originally may have been coming out sooner than it seems is now the case? Could be games made using those rumoured new devkits.
it's possible. though I can see them coming out for switch first and then Drake later, still. if they are being held for Drake, they're probably based on the switch versions rather than the xbox/playstation versions
ShadowFox08
Paratroopa
By 4x more RAM, i meant RAM bandwidth when I posted this, not actual RAM. That's my bad.
I have no idea what you're talking about. I was talking about the actual soc being brand new and a generational leap to someone who still believes in Switch Pro.
Raccoon
Fox Brigade
- Pronouns
- He/Him
okay but try to see the forest for the trees here
the comparison here is the line between wii u and switch, which needed a huge transition. regardless of what they call it, the new device will not immediately displace switch
ShadowFox08
Paratroopa
I never said Drake will immediately displace Switch. I argued that it's a new console with a generational leap in power, and which is why I said i'm not going into semantics about how Nintendo will advertise it at launch. Nobody knows exactly how it will be advertised, and considering Switch is still selling well, I don't think anyone here thinks it will be replaced, or at least not for a while.
I might have sounded harsh, but last time I checked, he still believes Drake will be a Pro model, which is something I don't want to go into further. He is entitled to his opinion.
Everyone here assumes it's going to be a hybrid and will keep the form factor. But the console is a proper generational leap in power with new features ,and who knows what kind of new tech/gimmick Drake will have that wasn't on previous consoles that isn't known already (DLSS, RT).
Last edited:
Cackletta5
Cappy
- Pronouns
- she/her
it’s funny how insiders gave us the time frame ‘second half 2022 or first half 2023’ and IF it’s even coming next spring, it might be for Zelda, so the very last moment of that time frame. But maybe even later. Guess we can’t just have nice things
- Pronouns
- he / him
Watching the video I see what a difference it makes, but as someone who hasn't been following along every version is there a simple reason as to why trees were distinct from other items that were nanite-compatible?
In the same way that 8:46PM is the very last moment of a day.
Stinky Horse
Bob-omb
In the original Nanite there were no options for opacity or animation. I guess Nanite based curtains would benefit in the same way too but nobody cares about curtains.
Raccoon
Fox Brigade
- Pronouns
- He/Him
from my perspective, you interjected into a product semantics argument with insistence that the silicon is next gen. I don't disagree, but in a discussion of what this product needs to sell its next gen-ness is irrelevant.
does drake need a big game to sell? if you don't think so, your counterpoint is a non sequitur
edit:
as for the idea that it could be something not-a-switch and therefore be a similar transition, sure. that's not a discussion of teraflops though
theoretically, there's no reason it can't work. but it's more likely a computational cost. recalculating all the connected edges and face deformation would throw things out of wack and cause unnecessary culling due to triangles changing size and shape. mesh shaders could handle trees better from what I hear, but would be slower than nanite if the triangles are smaller than a pixel, which would be likely if you went all out and modeled the leaves
Sol
Moblin
At this point I'm hoping Switch 2 comes out in the spring so that Pokemon Gen 10 will be exclusive to Drake.
Maybe then we can finally get a Pokemon game with a stable 30 FPS.
I have to take breaks playing Violet because there is so much stutter and frame drops that it hurts my eyes.
Maybe then we can finally get a Pokemon game with a stable 30 FPS.
I have to take breaks playing Violet because there is so much stutter and frame drops that it hurts my eyes.
ShadowFox08
Paratroopa
Alright, I'll bite. It's hard to say if Drake needs Botw 2 to sell in general. But if we are talking about a $400-500 console, and if it's not going to replace Switch right away... Nintendo would absolutely benefit A LOT from releasing Drake within Totk's launch window and start off with a strong momentum to justify the price.. That or another big first party game.
Everyone knows the switch's hardware has eally aged. Not just 3rd parties, but 1st party games with 600-720p docked games. We'll see how Totk performs. I can bet Nintendo will showcase the Drake and use Totk to sell it with improved performance, resolution and detail. Existing titles like Xenoblade 3, Bayonetta 3, Pokemon S/V are all good candidates. Not to mention Pikmin 4, FE (releasing in January).
LiC
Member
But that doesn't have anything to do with the resolution.
brainchild
Moblin
Oh for sure, I wasn't arguing otherwise. My point of contention was with the notion that it didn't take them too long to adapt to the industry standard at that time. They're a different company now, and how they behaved in the past will become less relevant to how they'll behave in the future over time. Whether that change makes it more likely that they'll quickly adapt to new industry standards is anyone's guess.
Honestly? I have been a bit humbled by my journey through this game development process. This is my first major solo indie project and it's proving to be a lot more difficult to handle than I was expecting, primarily due to me now focusing on turning this into a marketable product and not just a tech demo.
I don't want my game to be remembered for just having a cool gimmick or notable technical accomplishment; I want it to actually be a good game, so I'm now going through the painful process of reworking content that just doesn't seem fun or appealing, or may negatively impact the user experience. I am also considering how I will offer support for technical issues that may arise once the game releases. For those and other reasons, I have decided to intentionally slow down the development process to allow myself the necessary time to make a good product (and aside from my work as a Data Scientist, I also work full-time as an IT professional for a Fortune
I am currently not ready to release any promotional materials, as I want that aspect of the marketing to be handled as professionally as possible and I'm not in the position to do that right now. Also, it may be better to keep the timeline to reveal a secret until it's revealed, but I still haven't completely decided on that yet.
Last edited:
D
Deleted member 887
Guest
It doesn't apply to the Breath of the Wild case which is limited by memory bandwidth more than CPU or GPU, but frame time is frame time. Running physics calculations on a hundred on-screen objects at the same time, all of which have physics interactions with each other can leave you with very little time to render the scene. Is resolution the first place to find that time? No, probably not during development. But you'll certainly see DRS drops due to physics on systems that are CPU bound.
ReddDreadtheLead
#TeamLate2025WithAPotentialForEarly2026
- Pronouns
- He/Him
I’m serious about the Displaced Micro Mesh Engine and Opacity Micromap Engine…
It’s added per RT core it seems and makes the process more efficient… if there is anything backported from Ada that is.
It’s added per RT core it seems and makes the process more efficient… if there is anything backported from Ada that is.
Cuzizkool
Shriekbat
- Pronouns
- He/Him
A little off topic, but did anyone else read about Nvidia Thor? I’m having trouble deciphering the specs, but it states that it’s 2 PFLOPs. I read the fine print and it said (FP8). Does that mean that (at max wattage of course) it’s 500 TFLOPs (FP32)? Because that would be crazy if true. But I often get my TOPs and TFLOPS mixed up, so someone feel free to burst my bubble.
it's a big chip is why
Cuzizkool
Shriekbat
- Pronouns
- He/Him
So, this thing probably won’t have AGX or NX or Nano versions then? If it’s whole point is to provide raw centralized power for the cars it’s in, then can we definitively say that there can’t be a version of this chip dedicated to video game consoles?
there will probably be an AGX/NX/Nano version. those aren't size constrained devices. if there's a version dedicated to gaming, it'd be semi-custom like Drake is. Drake isn't a cut down Orin, it's a new chip that happens to use a similar CPU and GPU
- Pronouns
- He/Him
Yeah, it's 2000 TOPS (FP8) for deep learning. But that is a HUGE jump over Orin's ~275 TOPS maximum in the same category. But we don't know how that relates to TFLOPs, if it does at all (no small amount of that boost could be due to massive improvements to NVDLA).
Last edited:
Cuzizkool
Shriekbat
- Pronouns
- He/Him
See that’s so confusing to me. INT8, FP8, FP16, FP32… I’ll never get it. So I thought I’d bring it up here.
Thanks for filling me in. It will definitely be a powerful chip, no matter what. If Drake does turn out to be a Pro, and Switch 2 uses similar architecture from Thor, that would be quite a leap, for sure.
D
Deleted member 887
Guest
I doubt that there will be a Thor Nintendo console, merely because Thor is targeting a 2025 date, which seems early. That's only 2-3 years for Drake.
That said, Thor almost definitely will power a game console. Thor is designed to be able to drive dual screen infotainment engines, so GPS in the front of the car, games and streaming movies in the second screen in the back. And the interesting game viable tech in Thor will certainly wind up in any future SOC that Nvidia designs for Nintendo - if for no other reason than Thor's GPU is stock Lovelace.
The bigger question about Thor is Grace, the CPU inside Thor. I actually think there are real interesting possibilities to a coherent memory model for game consoles, but I have been unimpressed with the Neoverse cores.
ShadowFox08
Paratroopa
Who knows, but if you're wondering if a Switch successor down the line will use it in the near future, it won't. It's too big and draws too much power for a hybrid device, and as oldpuck says, it's a 2024/5 release. I guess it could be used in some cars for gaming though, like Nvidia 3000 series for the latest Tesla cars. But it doesn't seem like Nintendo is interested in a docked only console. Certainly not as the main draw. And if they did release one after Drake, it might be a modified Orion AGX at 5nm that is part of the Drake family and that would be more powerful than series s, but still weaker than PS5/Series S. Kinda of like a Ps2 vs Xbox. Not holding my breath on it though.
Tegra x1 to Tegra Orion/Drake (which uses Ampere) is 3 generations apart. If Drake ends up replacing switch eventually, then whatever is like 3 generations after Orion, might be a candidate.
Last edited:
Anatole
Octorok
- Pronouns
- He/Him/His
FP8 would still be TFLOPS; it’s different from INT8, which would be TOPS. FP8 was only made public a couple months ago (https://developer.nvidia.com/blog/n...ndardization-as-an-interchange-format-for-ai/).See that’s so confusing to me. INT8, FP8, FP16, FP32… I’ll never get it. So I thought I’d bring it up here.
OP
OP
Thor's GPU's probably going to deviate from consumer Ada Lovelace GPUs to at least some degree, similar to how Orin's GPU deviated from consumer Ampere GPUs to a degree.
And Nvidia mentioned Thor's using Arm Neoverse Poseidon AE for the CPU, the successor to the Neoverse V2, which Nvidia used for Grace.
Cuzizkool
Shriekbat
- Pronouns
- He/Him
How would FP8 flops convert to FP32 flops? Same as 16 to 32? 2000 fp8 tflops meaning 500 fp32 TFLOPs?
Anatole
Octorok
- Pronouns
- He/Him/His
That would be the goal in theory. I don’t think there are any released hardware implementations of FP8 to confirm this yet. In the AMD/Nvidia/Intel white paper, they emulate FP8 in a higher precision rather than using a specific hardware implementation.
EDIT: looked for the Lovelace whitepaper after seeing Oldpuck’s post, and I was wrong about this. FP8 is implemented in the Lovelace tensor cores.
For the RTX 4090 without accounting for sparsity acceleration (Appendix A, Table 2 of the whitepaper), the peak rates are:
FP16 accumulated to FP32: 165.2 tensor TFLOPS
FP16 accumulated to FP16: 330.3 tensor TFLOPS
FP8 accumulated to either FP16 or FP32: 660.6 tensor TFLOPS
Last edited:
D
Deleted member 887
Guest
The ratios of various floating point operations to each other and to the clock can vary from arch to arch. I'm not sure what they are for Ada
- Pronouns
- He/Him
OK, yeah, should have read more closely, conflated FP8 with INT8.
So it seems it's still talking about deep learning performance specifically, but moving data from floating point to integer seems to mean a loss in performance, so the problem is resolved going from FP to FP?
But I still don't think that can tell us anything else about what Thor is capable of.
And it's also kinda interesting to note that, given how Nvidia is iterating the Tegra line every 2 years or so, Thor might not be what gets considered for the hardware following Drake, as Thor is set for availability in 2025 unless it meets a delay and, if Drake releases in 2023, would only be 2 years into its hardware cycle. We'd maybe be looking at the follow-up to Thor in 2027/2028 for the basis of the next custom SoC design for Nintendo or something wholly-custom and highly-specialized to task entirely, to keep out of the current money pit that smaller processes than 6nm seem to be (because when Apple is locking more efficient processes for their silicon behind the Pro line for the first time ever, you know something's wrong).
Mr Swine
Like Like
- Pronouns
- He
Wouldn’t that have to be back ported at the same time as Ada and Drake were under development? seems a bit far fetched that it would be on Drake
mariodk18
Install Base Forum Namer
Assuming Nintendo sticks with their six year hardware lifespans gaps, I would expect the Nintendo Switch 3 to launch either in holiday 2028 or first half of 2029. What kind of hardware would be available around that time? Would it still be Thor or the next thing?
Last edited:
YolkFolk
Tingle
Series X level. In terms of raw power, it’s usually a handheld version of whatever the previous console generation was. It’ll have the benefit of the latest upscaling tech on top of that.
ReddDreadtheLead
#TeamLate2025WithAPotentialForEarly2026
- Pronouns
- He/Him
The Tegra X1 got features from later architectures while both were being developed. Ampere and Ada are similar, not the same just similar.
I don’t think it’s that far fetched.
I mean it’s not like it’s offering a performance uplift that necessitates the whole line getting the upgrade. And they recognize that their gains have diminished over the years. If they do it, it’s for other reasons.
Apple is also locking down faster USB speeds to the Pro models.
It’s what I’d call, “BoM optimization” because they can.
D
Deleted member 1324
Guest
If Tears of the Kingdom is 1920x1080 in any way, shape or form including dynamic on the launch Switch I will gladly buy you a copy of the game.
Nintendo are likely really pushing Switch to breaking point with this game in terms of it’s rendering quality of material etc so choosing to render the game at ~50% more pixels at the low end versus BotW (considering BotW actually drops to 720p at times) seems completely unrealistic to me.
TotK will run at dynamic 900p when docked on Switch and spend much of its time below that number imo. It will still have framerate drops below 30fps.
The Drake version will be 2160p/60fps using DLSS when docked in comparison.
Anatole
Octorok
- Pronouns
- He/Him/His
Yeah, it’s different in a couple of ways. You don’t actually train a network in INT8; instead, you train it in FP16 or FP32, then quantize to INT8 for inference. This allows you to take advantage of the much larger dynamic range and non-uniform distribution (some half of floating point values are in the range [-1,1]) of floats during training, which is especially important for representing gradients. Once you quantize to ints, you introduce some error from clipping at the ends of the dynamic range and rounding error within the range, but there are ways to minimize this error. Nvidia has a nice blog post about it, with a good summary graphic of quantization:
One of the goals of FP8 is to allow you to do not just inference in 8-bit precision, but also training. The big design question for FP8 is how many bits should be dedicated to the order of magnitude (the exponent) and how many to the number itself (the mantissa). It’s basically a tradeoff between dynamic range and precision. The Nvidia/ARM/Intel white paper for FP8 actually uses two different representations: E4M3 (four bits for the exponent, 3 for the mantissa) for tensors representing weights/activations and E5M2 for tensors representing gradients. They show in section 4 of the white paper that many common ML models can be trained in FP8 with a nearly identical accuracy to FP16.
D
Deleted member 1324
Guest
What’s your point?
The sequel to Breath of the Wild will arguably be Nintendo’s most hyped up game ever. Their ‘Halo 2’. We know from Nvidia leaks they have an incredibly powerful SoC. We know from Bloomberg that dev kits have been out for over a year. We can speculate that a combination of Covid and chip shortages have potentially delayed the console from a planned late 2022 release into 2023. What game comes out in 2023 that’s big enough to support the launch of a new console?
Nintendo would be quite literally stupid not to launch the console with Tears of the Kingdom next May.
ReddDreadtheLead
#TeamLate2025WithAPotentialForEarly2026
- Pronouns
- He/Him
I’m looking at the proposed specs, and I question if they even need DLSS to resolve a 4k60 image.
If they start pushing settings higher sure, but I don’t really think DLSS is needed to take a 1080p switch game to 4K.
It’s… a switch game… it wouldn’t be that different from XBox One X to XBox One games… that were 4K on the former vs whatever the XBox did.
If they start pushing settings higher sure, but I don’t really think DLSS is needed to take a 1080p switch game to 4K.
It’s… a switch game… it wouldn’t be that different from XBox One X to XBox One games… that were 4K on the former vs whatever the XBox did.
Makes me wonder if they’ll rely more on FP8 for DLSS while seemingly speeding up the process without any or much loss in quality.Yeah, it’s different in a couple of ways. You don’t actually train a network in INT8; instead, you train it in FP16 or FP32, then quantize to INT8 for inference. This allows you to take advantage of the much larger dynamic range and non-uniform distribution (some half of floating point values are in the range [-1,1]) of floats during training, which is especially important for representing gradients. Once you quantize to ints, you introduce some error from clipping at the ends of the dynamic range and rounding error within the range, but there are ways to minimize this error. Nvidia has a nice blog post about it, with a good summary graphic of quantization:
One of the goals of FP8 is to allow you to do not just inference in 8-bit precision, but also training. The big design question for FP8 is how many bits should be dedicated to the order of magnitude (the exponent) and how many to the number itself (the mantissa). It’s basically a tradeoff between dynamic range and precision. The Nvidia/ARM/Intel white paper for FP8 actually uses two different representations: E4M3 (four bits for the exponent, 3 for the mantissa) for tensors representing weights/activations and E5M2 for tensors representing gradients. They show in section 4 of the white paper that many common ML models can be trained in FP8 with a nearly identical accuracy to FP16.
I want my ray traced shadows and AO!
ReddDreadtheLead
#TeamLate2025WithAPotentialForEarly2026
- Pronouns
- He/Him
RGTI and I’ll accept the DLSS being needed
I think it's obvious they will push settings, at least I could see Draw distance getting a huge boost.
It's a game about verticality. On a huge 4K screen, you want to soak in all that beautiful landscape detail, while falling thousands of feet above ground.
Also, pushing settings form a development perspective is probably the cheapest way to get a visual boost.
Last edited:
