Dekuman
Kremling
- Pronouns
- He/Him/His
That would be nice but nature abhors a vacuum of information, as long as this thread exists it will get filledAlternative approach: everybody can just chillax.
That would be nice but nature abhors a vacuum of information, as long as this thread exists it will get filledAlternative approach: everybody can just chillax.
Sure smells like “we can now say with certainty that the new hardware is a powered up revision that is fully back compat, and Switch is still the platform for the foreseeable future”
There's plenty of interesting stuff in the leak (not all related to Drake) that's gotten zero discussion because no one has looked at it. But oh well. The tea leaf reading will continue.That would be nice but nature abhors a vacuum of information, as long as this thread exists it will get filled
I've avoided the actual dump of data, partially to avoid encouraging that sort of behavior (the dump of employee records was especially abominable), but I keep being tempted to pick it up and peek throughThere's plenty of interesting stuff in the leak (not all related to Drake) that's gotten zero discussion because no one has looked at it. But oh well. The tea leaf reading will continue.
I don't think it's right to redistribute proprietary source code, or disclose details of sensitive implementations thereof, but looking through it once it's out there is merely akin to data mining if you're not interested in stealing and sharing trade secrets. But I get it.I've avoided the actual dump of data, partially to avoid encouraging that sort of behavior (the dump of employee records was especially abominable), but I keep being tempted to pick it up and peak through
WHAT?!Alternative approach: everybody can just chillax.
On a more serious note, maybe the reason for that is that in this thread most people that are tech-inclined (including you) and could look into that info and make sense out of it, are trying to not derail it with non-“Future Nintendo Hardware” stuff. And other people like me just follow along.There's plenty of interesting stuff in the leak (not all related to Drake) that's gotten zero discussion because no one has looked at it.
I think this confirms they have no insider info, because if they had then they wouldn’t post about a “Nintendo Next Gen” while under NDA.I backed the game. They sent this out after the kickstarters end.
Alternative approach: everybody can just chillax.
CPU setup | L1 Cache | L2 cache | L3 Cache | Total Range L1 | Total Range L2 | Total Range L3 |
8 A78C | 32kB to 64kB of L1 Instruction Cache + 32kB to 64kB of L1 Data Cache per core | 256kB to 512kB L2 private ECC per core | up to 8MB L3 (shared) | 256KB to 512KB L1 I-Cache + 256KB to 512kB L1 D-Cache = 512kB to 1MB L1 Cache | 2-4MB L2 Cache | up to 8MB L3 (shared) |
6 A78C | 32kB to 64kB of L1 Instruction Cache + 32kB to 64kB of L1 Data Cache per core | 256kB to 512kB L2 private ECC per core | up to 6 or 8MB L3 (shared) | 192kB to 384kB L1 I-Cache + 192kB to 384kB L1 D-Cache = 384kB to 768kB L1 Cache | 1.5-3MB L2 Cache | up to 6 or 8MB L3 (shared) |
4 A78 | 32kB to 64kB of L1 Instruction Cache + 32kB to 64kB of L1 Data Cache per core | 256kB to 512kB L2 private ECC per core | 512kB-4MB L3 (shared) | 128kB to 256kB L1 I-Cache + 128kB to 256kB L1 D-Cache = 256kB to 512kB L1 Cache | 1-2MB L2 Cache | 512kB to 4MB L3 (shared) |
4 A57 | 48kB Instruction Cache w/DED Parity + 32KB Data Cache w/ECC per core | 512kB to 2MB L2 Cache (shared) | N/A | 192kB L1 I-Cache + 128kB L1 D-Cache = 320kB L1 Cache | 512kB to 2MB L2 Cache | N/A |
4 Jaguar Cores | 32kB of L1 Instruction Cache + 32kB of L1 Data Cache per core | 1 to 2MB L2 Cache (shared) | N/A | 8 Core custom contains 512kB L1 cache (256kB in 4 Core Cluster) | 8 core custom contains 4MB L2 Cache (2MB in 4 Core Cluster) | N/A |
Likely not, Switch released during the time that Pascal was out and that was noticeably better than the Maxwell 2.0.@NateDrake
Is there a chance that the delay from 2021 to X is indeed because of Ada? Ada seems to be a massive step up compared to Ampere and even RDNA 3 according to the last rumors from greymon and kopite…
Obviously, this doesn’t require a tensor core to execute but is another answer to the question “what else can tensor cores do for games?”In this title, how did you overcome what seemed technically impossible at the time?
Okane: We just did a lot of trial and error. In terms of new technology, we tried deep learning, which was not available when Wii Sports was developed.
Deep learning?
Okane: For example, we asked various people to "swing as if you were hitting to the left," and "now swing as if you were hitting to the right," and repeated the process over and over again to acquire a large amount of data. It is a technology that detects if the player is trying to hit the ball to the right (or left) by statistically analyzing that acquired data. So by incorporating such technology, and experimenting with this and that, we tried to bring it up to a level that everyone could agree on.
that's one of those things I've been wondering myself. especially if Nvidia wants to further justify their inclusion in geforce cards. I've heard that fp16 tasks have ran on tensor cores but nothing really consistent on that. if anyone has any motivation for find use for silicon, Nintendo would since any solution would probably be bespoke to the hardwareNintendo used Deep Learning to develop the motion controls for Switch Sports
Ask the Developer Vol. 5, Nintendo Switch Sports–Part 2 - News - Nintendo Official Site
Content pre-recorded in accordance with current COVID-19 health and safety guidelines. This article has been translated from the original Japanese content. This interview was conducted before the game was released. Contents 1: Aiming for the world’s…www.nintendo.com
Obviously, this doesn’t require a tensor core to execute but is another answer to the question “what else can tensor cores do for games?”
I suspect that offline voice assistants and ultra smooth scrolling UHD devices are going to put matrix math acceleration everywhere, regardless.that's one of those things I've been wondering myself. especially if Nvidia wants to further justify their inclusion in geforce cards. I've heard that fp16 tasks have ran on tensor cores but nothing really consistent on that. if anyone has any motivation for find use for silicon, Nintendo would since any solution would probably be bespoke to the hardware
On top of the existence of the OLED, I don't think there's any version of Drake that could have released in 2021 given the state of NVN2 and other things as of the February 2022 leak. Later this year is the earliest it was ever going to come out. As for the Ada thing, there's no evidence that it's any more like Ada than the original Switch is like Pascal -- potentially a few late feature inclusions, but no significant changes to the architecture.@NateDrake
Is there a chance that the delay from 2021 to X is indeed because of Ada? Ada seems to be a massive step up compared to Ampere and even RDNA 3 according to the last rumors from greymon and kopite…
Don't know what you are talking about a supposed delay, but based on process node alone 5nm tmsc is a massive improvement over Sammy 8nm. So even Ampere at 5nm would have been a massive improvement.@NateDrake
Is there a chance that the delay from 2021 to X is indeed because of Ada? Ada seems to be a massive step up compared to Ampere and even RDNA 3 according to the last rumors from greymon and kopite…
I mean, machine learning is just the usual way to do stuff like that today. tensor cores are great for that, sure, but for stuff like motion clasification you dont have that many features, and it mostly is done prior to deployment. Aftere that, the model is there, you ship it with your software, and use that to clasify the movement.Nintendo used Deep Learning to develop the motion controls for Switch Sports
Ask the Developer Vol. 5, Nintendo Switch Sports–Part 2 - News - Nintendo Official Site
Content pre-recorded in accordance with current COVID-19 health and safety guidelines. This article has been translated from the original Japanese content. This interview was conducted before the game was released. Contents 1: Aiming for the world’s…www.nintendo.com
Obviously, this doesn’t require a tensor core to execute but is another answer to the question “what else can tensor cores do for games?”
Absolutely. I wasn't expecting that you would train the model on the console. You'd want something like a DLA for that anyway - but model execution could run on tensor cores. My point was just that being able to accelerate DL models will benefit more than just the graphics pipeline. Even if each individual model is very fast on the CPU, once they're everywhere, they will consume a significant portion of the frame budgetI mean, machine learning is just the usual way to do stuff like that today. tensor cores are great for that, shure, but for stuff like motion clasification you dont have that many features, and it mostly is done prior to deployment. Aftere that, the model is there, you ship it with your software, and use that to clasify the movement.
Later in the interview it is mentioned that netplay uses a different motion classification method than local play, because they can't wait for all the motion inputs to be collected before they need to respond. Rollback netcode is built on the idea that you predict player actions, and then roll back when you have updated state. Player action prediction is exactly the sort of thing that tensor cores could accelerate.If im wrong im happy to be corrected. And if somebody has an idea where else besides the input could be used (i mean, technically to train AI to your playstyle, see Figure Playeres in Smash 4/Ultimate...), i would love to hear it.
In short: Technically possible, but for that you dont need the tensor cores. Graphics (and maybe some complexe audio stuff) are the only areas where i see benefits.
Yes to all of those. Just one question: how much of those problem spaces do you have at once in a game?Absolutely. I wasn't expecting that you would train the model on the console. You'd want something like a DLA for that anyway - but model execution could run on tensor cores. My point was just that being able to accelerate DL models will benefit more than just the graphics pipeline. Even if each individual model is very fast on the CPU, once they're everywhere, they will consume a significant portion of the frame budget
Later in the interview it is mentioned that netplay uses a different motion classification method than local play, because they can't wait for all the motion inputs to be collected before they need to respond. Rollback netcode is built on the idea that you predict player actions, and then roll back when you have updated state. Player action prediction is exactly the sort of thing that tensor cores could accelerate.
You would likely want to use a model that takes player's previous actions as inputs, but that isn't training.
AI pathfinding is a problem made for deep learning. As good as modern pathfinding is (and it's mostly pretty good!), it bogs down as enemy numbers increase and still depends on navigation meshes (even if those meshes are automatically generated from scene geometry).
that reminds me, remember Hyperscape? turns out it was good for something!AI pathfinding is a problem made for deep learning. As good as modern pathfinding is (and it's mostly pretty good!), it bogs down as enemy numbers increase and still depends on navigation meshes (even if those meshes are automatically generated from scene geometry).
Nope just some people still trying to come to terms as to why new powerful hardware didn’t show up in 2021, maybe this year too.wait.. Was there some new info about a delay?
But you and me both admitted to becoming zombified corn on * checks notes *….July 6, 2021"it's not fake! it's not fake!" I continue to insist as I slowly shrink and transform into a corn cob
When you have a 100m + install base it becomes more about selling software than hardware for Nintendo profits.As much as I want Switch 2 to be released Q4 of this year, Nintendo really has no need to, and it might end up hurting their overall switch 1 sales. They can survive from Splatoon 2 and Pokemon (and any unannounced games) during the holidays.
If they want to maximize current switch model sales, it's best to reveal Switch 2/Drske next year in January, if they are aiming for a spring release with Botw 2.
In the short term I expect it to be a niche solution. Medium term I expect this to be everywhere mostly because it will be built by default into engines. Unity/Epic have the resources to build models for lots these tasks, and hide them entirely from devs - it won't be about building custom engines that use tensor cores to accelerate game play, it will be around using the hardware to accelerate common tasks. Switch Sports is an example of DL model that is embedded in a total banal part of the game - we're not going to be able to anticipate all the places devs will find uses for it.Yes to all of those. Just one question: how much of those problem spaces do you have at once in a game?
I just feel like you would need to force it to get to a place where Tensor Cores really make a tangible difference for gameplay.
I dont see it in mainstream games at all. Some niche genres, experimental games, simulation based genres...yes.
what they mean is inference. they train the model offline and then the game refers to the model in-game. the issue here is that it doesn't take a lot of resources to refer to the model. hence the lack of need for tensor coresI thought it was just Switch Sports's development that used machine learning. Surely it doesn't have any in-game?
ah interesting, I thought they had just used the model to set parameterswhat they mean is inference. they train the model offline and then the game refers to the model in-game. the issue here is that it doesn't take a lot of resources to refer to the model. hence the lack of need for tensor cores
By necessity, hardware tends to arrive more or less when it's ready, not at the perfectly minmaxed point in time for it to arrive.As much as I want Switch 2 to be released Q4 of this year, Nintendo really has no need to, and it might end up hurting their overall switch 1 sales. They can survive from Splatoon 2 and Pokemon (and any unannounced games) during the holidays.
If they want to maximize current switch model sales, it's best to reveal Switch 2/Drske next year in January, if they are aiming for a spring release with Botw 2.
you don't need Orin for that. any computer would doNintendo could, in theory, use a tegra Orin developer kits and train different neural networks for different use cases and implement it into different games. Like movement that would function similar to the Wii’s movement input.
Or fine tune the gyro
Or what have you.
By necessity, hardware tends to arrive more or less when it's ready, not at the perfectly minmaxed point in time for it to arrive.
Besides, while they might get more short term profit from people buying Switch 1, building up a Switch 2 install base would be better long term.
I'm not sure of it, but late console buyers probably don't buy a whole lot of software
How funny would it be if we get a ps5 pro or the next Xbox before the next switch.
Yeah me too.I feel like it's got a good chance of happening unfortunately. Nintendo ain't wanting to budge at all on new hardware lol.
Do you think Nintendo thinks it's ready?By necessity, hardware tends to arrive more or less when it's ready, not at the perfectly minmaxed point in time for it to arrive.
Besides, while they might get more short term profit from people buying Switch 1, building up a Switch 2 install base would be better long term.
I don't think it will happen before then. With the chip shortages, it's hard enough as it is finding a PS5 and x series x. The X series x/s and PS5 are currently close to the 15 and 20 million sales mark worldwide. X series x/s actually outsold PS5 by a few thousand in the U.S. Both were above 7 million as of April. Helps that S is much more readily available and $300.How funny would it be if we get a ps5 pro or the next Xbox before the next switch.
Not funnyHow funny would it be if we get a ps5 pro or the next Xbox before the next switch.
When the hardware will be ready is largely dictated by things scheduled pretty far in advance. They can only hold it back so long. If the rumors so far are accurate, next Spring is probably a fairly hard limit.Do you think Nintendo thinks it's ready?
I don't think Nintendo will wait too long for a successor release.
I think there's a higher chance if it releasing with botw 2 vs this fall/winter 2023. By Q4 2023 is not off the cards either, but I hope they don't wait longer than spring..
My point was that it can be trained on something so small. A massive super computer isn’t needed.you don't need Orin for that. any computer would do
Before I'm not sure, but within the same year, it's absolutely possible as I don't see the next switch coming out before 2024, which may be the right time for some pro consoles.How funny would it be if we get a ps5 pro or the next Xbox before the next switch.
Uhhh what exactly are you basing this on?I feel like it's got a good chance of happening unfortunately. Nintendo ain't wanting to budge at all on new hardware lol.