Fingers crossed that these RT cores end up in Dane.
Edit: Woops, I totally misread your comment. Considering Nvidia has made no mention of which generation the RT cores on Orin are part of, there's definitely a possibility.
Last edited:
Fingers crossed that these RT cores end up in Dane.
He meant “these” as in gen 3 rt cores.Assuming the rumour that Dane's a customised version of Orin is true, physically removing the RT cores from the GPU would be cost prohibitive for Nintendo and Nvidia. Also, NateDrake mentioned that late 2020 devkits have limited RTX support; and I imagine there's no reason to include limited RTX support if there was no plans to use RT cores.
Thank you for letting me know. I can't believe I totally misread the comment.He meant “these” as in gen 3 rt cores.
Well, even if they aren’t, Dane IS supposed to be a custom chip this time, allegedly with Lovelace features, so this could be one of the Lovelace features being talked about.Assuming the rumour that Dane's a customised version of Orin is true, physically removing the RT cores from the GPU would be cost prohibitive for Nintendo and Nvidia. Also, NateDrake mentioned that late 2020 devkits have limited RTX support; and I imagine there's no reason to include limited RTX support if there was no plans to use RT cores.
Edit: Woops, I totally misread your comment. Considering Nvidia has made no mention of which generation the RT cores on Orin are part of, there's definitely a possibility.
I find it unlikely that Dane will deviate that much from Orin.It will be through and through the same generation of chip imo.Well, even if they aren’t, Dane IS supposed to be a custom chip this time, allegedly with Lovelace features, so this could be one of the Lovelace features being talked about.
No one knows how custom Dane is. Dane could be custom in the sense that the Cortex-A78C is used instead of the Cortex-A78AE; and all of the hardware features that aren't useful for game development, such as a safety island, programmable vision accelerators (PVA), etc., are removed.Well, even if they aren’t, Dane IS supposed to be a custom chip this time, allegedly with Lovelace features, so this could be one of the Lovelace features being talked about.
My understanding is that Nvidia customized the TX1 beyond the standard Maxwell config to include Pascal features. And that was for their own use in the Shield that the TX1 was seemingly designed for. If they’ll do that kind of SoC config for themselves, I can’t see a reason they wouldn’t provide the same courtesy to Nintendo, given how big of a customer they’re going to be.I find it unlikely that Dane will deviate that much from Orin.It will be through and through the same generation of chip imo.
Well, whether or not it goes past that is going to be dependent on whether kopite was correct in saying that Dane will have Lovelace features, which doesn’t involve CPU changes. It’s also my understanding that he was saying that about Dane specifically rather than Orin overall.No one knows how custom Dane is. Dane could be custom in the sense that the Cortex-A78C is used instead of the Cortex-A78AE; and all of the hardware features that aren't useful for game development, such as a safety island, programmable vision accelerators (PVA), etc., are removed.
I was simply stating a couple of examples.Well, whether or not it goes past that is going to be dependent on whether kopite was correct in saying that Dane will have Lovelace features, which doesn’t involve CPU changes. It’s also my understanding that he was saying that about Dane specifically rather than Orin overall.
Ahh, so he was saying it’s Orin overall that borrows a bit from Lovelace, my bad. I guess we’ll see what gen of RT cores the Orin configurations come with and then figure it from there.I was simply stating a couple of examples.
Anyway, so far, kopite7kimi seems to be correct about the Orin's GPU being based on Lovelace, considering Orin's GPU has 50% more L1 cache in comparison to GA102 (192 KB for Orin's GPU vs 128 KB for GA102), and Orin's GPU supports AV1 encoding. (kopite7kimi did mention that Lovelace has larger cache, and consumer Ampere GPUs only supported AV1 decoding.)
Yes they should be able to. Nintendo has done it before with CPU cores. The A53 cores on the TX1 Switch were disabled since A53s and A57 cores couldn't be run at the same time, so it felt like a waste (and to save energy), so one A57 core was dedicated for the OS.If devs don’t use the RT cores, can they be shut off manually/automatically to save power?
I do believe the tx1 featureset is the same for every version of the chip.My understanding is that Nvidia customized the TX1 beyond the standard Maxwell config to include Pascal features. And that was for their own use in the Shield that the TX1 was seemingly designed for. If they’ll do that kind of SoC config for themselves, I can’t see a reason they wouldn’t provide the same courtesy to Nintendo, given how big of a customer they’re going to be.
Can't really "turn them off", through in a practical sense, not using them means they are not active. That said, if they were unusable, they'd be fused off, or outright not be thereIf devs don’t use the RT cores, can they be shut off manually/automatically to save power?
I should note that the Cortex-A53 cores being disabled isn't limited to the Nintendo Switch. Any device running on the Tegra X1 is practically only running on the Cortex-A57 cores, considering Nvidia made no mention of the Cortex-A53 cores when talking about the specs of the Jetson TX1 (here and here); and the Cortex-A53 cores are also disabled in the Pixel C, which also runs on the Tegra X1.Yes they should be able to. Nintendo has done it before with CPU cores. The A53 cores on the TX1 Switch were disabled since A53s and A57 cores couldn't be run at the same time, so it felt like a waste (and to save energy), so one A57 core was dedicated for the OS.
Are the Cortex-A53s present in Mariko chips?I should note that the Cortex-A53 cores being disabled isn't limited to the Nintendo Switch. Any device running on the Tegra X1 is practically only running on the Cortex-A57 cores, considering Nvidia made no mention of the Cortex-A53 cores when talking about the specs of the Jetson TX1 (here and here); and the Cortex-A53 cores are also disabled in the Pixel C, which also runs on the Tegra X1.
yea, they're there.Are the Cortex-A53s present in Mariko chips?
Makes me wonder about the viability of the Cortex-A78C. Been wondering if there an even more powerful ARM CPU available out there, or if Nvidia could make one with custom (built-in x86/64) instructions such as the Apple M1,
A x86-64 to Arm binary translator (e.g. Rosetta 2) can hurt gaming performance, as shown with the Apple M1 Pro and the Apple M1 Max.Been wondering if there an even more powerful ARM CPU available out there, or if Nvidia could make one with custom (built-in x86/64) instructions such as the Apple M1,
The Cortex-A710 is also optimised for a 7 nm** process node as well, going by Cadence's press release and WikiChip's article on Matterhorn. So the Cortex-A710 can't be used if Dane's still being fabricated using Samsung's 8N process node, which I think is most likely.and the 710 is only for 5nm, I think
As mentioned, there's the A78's successor A710, which IIRC isn't that much better. And then there are the X1/X2 which are not fit for a Switch (they are made for small bursts or for non portable devices).Are the Cortex-A53s present in Mariko chips?
Makes me wonder about the viability of the Cortex-A78C. Been wondering if there an even more powerful ARM CPU available out there, or if Nvidia could make one with custom (built-in x86/64) instructions such as the Apple M1,
hell nah. that's way too expensive now. amortizing costs via using existing products is the only real path forward unless Nintendo wants to be on some really forward thinking designs. there are some posts on Beyond3D by an Epic Games engineer about how hardware needs to change how they do things since scaling is pretty much dead in the water. UE5 is trying to change that paradigm, but the hardware will just have to use work arounds until they are specifically designed around less hardware acceleration and smarter usage of computeI wonder if Nintendo is really interested in throwing money at Nvidia creating a custom design or they just use a “custom” design like they did with TX1.
Well, Nvidia has nothing that directly fits their needs, so it will be more "custom" than tx1.I wonder if Nintendo is really interested in throwing money at Nvidia creating a custom design or they just use a “custom” design like they did with TX1.
I think outside of changing the CPU from the Cortex-A78AE to probably the Cortex-A78C, and removing the hardware features that Nintendo doesn't need, such as the safety island, programmable vision accelerators (PVA), etc., probably the latter.I wonder if Nintendo is really interested in throwing money at Nvidia creating a custom design or they just use a “custom” design like they did with TX1.
They can be repurposed for sound/audio if Nintendo chooses. Fancy audio mind you.If devs don’t use the RT cores, can they be shut off manually/automatically to save power?
IIRC, they were fused off aka they physically aren’t there.Are the Cortex-A53s present in Mariko chips?
Makes me wonder about the viability of the Cortex-A78C. Been wondering if there an even more powerful ARM CPU available out there, or if Nvidia could make one with custom (built-in x86/64) instructions such as the Apple M1,
It’ll be as custom as the Series consoles and the PS5 APU are, taking existing designs and going from there to make them work together. Removing some features not necessary or adding from existing designs.I wonder if Nintendo is really interested in throwing money at Nvidia creating a custom design or they just use a “custom” design like they did with TX1.
I don’t really think Nintendo needs to do something like this anyway. But thank you for bringing it up.A x86-64 to Arm binary translator (e.g. Rosetta 2) can hurt gaming performance, as shown with the Apple M1 Pro and the Apple M1 Max.
I want to see Nintendo go whole hog on things like mesh shaders. I'm really curious as to what that could do for games
I feel safe saying the latter.I think outside of changing the CPU from the Cortex-A78AE to probably the Cortex-A78C, and removing the hardware features that Nintendo doesn't need, such as the safety island, programmable vision accelerators (PVA), etc., probably the latter.
yea, it can't be removed. it would solve some issues with assets being too low poly. like in Zelda, having terrain that has better detail up close would help a lotI don’t think Nintendo would even remove those, it’s integral to the uArch. So expect next Zelda to look noice.
and Luigi’s Mansion
And possibly Prime game.
yea, it can't be removed. it would solve some issues with assets being too low poly. like in Zelda, having terrain that has better detail up close would help a lot
Well yeah, emulating or translating is worse than native performance, but that's all irrelevant here...outside of talking about emulated stuff like running PPC Nintendo games (even then, I vaguely recall PPC emulation on ARM can be pretty quick anyway).A x86-64 to Arm binary translator (e.g. Rosetta 2) can hurt gaming performance, as shown with the Apple M1 Pro and the Apple M1 Max.
Yeah just treating "custom" as a black or white term is a bit misleading. There's anything from essentially mixing and matching from an existing SoC parts bin, to really custom like what Apple does for themselves. I'd expect the former with some further chopped down Orin part, that's still usable for Nvidia as well.It’ll be as custom as the Series consoles and the PS5 APU are, taking existing designs and going from there to make them work together. Removing some features not necessary or adding from existing designs.
The expense of going balls to the wall custom is so high it’s just not worth it. Unless you’re Apple or something.
I don't know if that's necessarily the case in reality, considering that Nvidia made no mention of the Cortex-A53 cores when talking about the Jetson TX1 specs (here and here); and the Cortex-A53 cores are completely disabled in the Pixel C, which is running on the Tegra X1, and would definitely benefit from having the Cortex-A53 cores available for use. And the Nintendo Switch dev kit spec sheet also made no mention of the Cortex-A53 cores.We have to keep in mind that the TX1 CPU is basically a 4 core CPU that happens to switch between 4*53 and 4*57 when clocks are under a certain load which is probably under the fixed 1GHz clock rate meaning that it's only running on the 4*A57.
We will never know but cluster switch should completely prevent the use of A53s. It would be a 4 core CPU even with the A53 activated.I don't know if that's necessarily the case in reality, considering that Nvidia made no mention of the Cortex-A53 cores when talking about the Jetson TX1 specs (here and here); and the Cortex-A53 cores are completely disabled in the Pixel C, which is running on the Tegra X1, and would definitely benefit from having the Cortex-A53 cores available for use. And the Nintendo Switch dev kit spec sheet also made no mention of the Cortex-A53 cores.
Those would most likely only be for OS related tasks, which they can just use A57 which offer the similar levels of perf to the A510 on a better node.Nintendo could also use 2*A510 or 4*A510 in a merged core configuration. That said A510 being 64bit limited could be a problem for 32bit games.
I think as far as released commercial mobile Arm based SoCs are concerned, Samsung's 14LPE process node is the most advanced process node used for the Cortex-A57, which is one of the CPU cores that the Exynos 7420 uses. But I think TSMC's 10FF process node is the most advanced process node used for the Cortex-A57, which is used for TSMC's 10FF validation SoC.An A57 is probably the same transistor budget as an A510, however I do not think A57 can be paired with the A78 or be on a node less than 7nm.
why would they do this?And there was a rumour about Nintendo requesting a Tegra X1 sample from Nvidia, where TSMC's 7 nm** process node is used for fabrication, and the frequency of the Cortex-A57 cores increased from 1.9 GHz to 2.52 GHz, despite Nvidia and TSMC warning Nintendo about the difficulty of exceeding the Cortex-A57's max frequency of 1.9 GHz. And the power consumption was apparently higher than expected.
Smaller process nodes = more SoCs fabricated per production run = cheaper productionwhy would they do this?
it seems weird
allows them to get a "Pro" device while easily keeping compatibility. all those fears of lack of BC would have never happened. but going to new architectures is the best option at the end of the day. they get more of everything, even if they have to put in a bit more work for BCwhy would they do this?
it seems weird
More like going to a new one was the only available option for a "pro" refresh, which ended up not only being a "pro", but a generational upgrade and using cutting edge technology at that.allows them to get a "Pro" device while easily keeping compatibility. all those fears of lack of BC would have never happened. but going to new architectures is the best option at the end of the day. they get more of everything, even if they have to put in a bit more work for BC
There was definitely a sense in the tests done around launch that the Switch wasn't really getting everything it could out of the microSD slot, as I believe there wasn't really a measurable difference in load times past a certain point. That said, those graphs seem to show a fair amount of game-to-game variance, and the tests around launch didn't exactly have a lot of choices, so it's possible that all the games tested were coincidentally games where the eMMC is faster for some reason. I'd be interested to see the other Switch models on those graphs.(I'm sharing the results of a recent storage speed test with the OLED Model. If this was already posted, let me know and I'll delete this.)
Numerous storage speed tests with the OG or Mariko models all indicated that internal eMMC > fast microSD > Game Card > slow microSD. Here is a typical outcome:
A Japanese website, Akiba PC Hotline, ran a series of tests comparing the load times of OLED Model eMMC to the latest generation of Samsung EVO Plus microSD, and the results are interesting.![]()
The microSD won 4 tests out of 6 by tiny margins, and eMMC won 2 by larger margins. This is rather unexpected, and I can think of the following potential explanations.1. ACNH Happy Home Paradise / 2. Apex Legends / 3. Fortnite / 4. Monster Hunter Rise / 5. Diablo II Resurrected / 6. NBA 2K22
![]()
![]()
![]()
![]()
![]()
![]()
On the one hand, this seems to suggest that there still is a bit headroom for microSD and eMMC, but on the other we probably can't expect a leap in performance unless Dane adopts a different technology, such as UFS or NVMe.
- OLED Model's 64GB eMMC is slower than the old 32GB eMMC: Improbable
- OLED Model reads microSD faster than older models: Not sure how to verify this until someone hacks the OLED Model
- Improved sequential transfer speed of EVO Plus: The previous gen was rated "up to 100MB/s", and the latest gen is "up to 130MB/s"
- Better random access speed of class A2: The new EVO Plus is rated A2, while most microSDs are only A1. Class A2 reads 1.67x faster and writes 3x faster than A1. Note that the host device needs to support the A2 standard to take advantage of the speed.
Huh...How long ago was this rumor?And there was a rumour about Nintendo requesting a Tegra X1 sample from Nvidia, where TSMC's 7 nm** process node is used for fabrication, and the frequency of the Cortex-A57 cores increased from 1.9 GHz to 2.52 GHz, despite Nvidia and TSMC warning Nintendo about the difficulty of exceeding the Cortex-A57's max frequency of 1.9 GHz. And the power consumption was apparently higher than expected.
** → a marketing nomenclature used by all foundry companies
It's fairly old. And, no, nvidia wouldn't be nor should Nintendo pay for it. They should just avoid it. BC isn't as big an issue as it seems. I think Nintendo just attempted the path of least resistance first (clock everything faster thanks to die shrinks) only for 2012 micro arch to rear its headHuh...How long ago was this rumor?
Nintendo must see something about the Tegra architecture, but I do wonder if Nvidia would be up to the task trying to create a fully back-compat SoC with "Dane" or whatever they have in store.
Oh, if it's an old rumor then definitely a lot has changed then.It's fairly old. And, no, nvidia wouldn't be nor should Nintendo pay for it. They should just avoid it. BC isn't as big an issue as it seems. I think Nintendo just attempted the path of least resistance first (clock everything faster thanks to die shrinks) only for 2012 micro arch to rear its head
You can get faster eMMC, sure, but you pay more money for it, likewise with UHS-II/UHS-III SD card readers and the cards themselves (there's not all that much to squeeze from UHS-I cards). And for the kind of price they'd pay to squeeze eMMC and microSD for all they're worth, options with far better performance and capacity per dollar start coming into play, including but not limited to UFS.(I'm sharing the results of a recent storage speed test with the OLED Model. If this was already posted, let me know and I'll delete this.)
Numerous storage speed tests with the OG or Mariko models all indicated that internal eMMC > fast microSD > Game Card > slow microSD. Here is a typical outcome:
A Japanese website, Akiba PC Hotline, ran a series of tests comparing the load times of OLED Model eMMC to the latest generation of Samsung EVO Plus microSD, and the results are interesting.![]()
The microSD won 4 tests out of 6 by tiny margins, and eMMC won 2 by larger margins. This is rather unexpected, and I can think of the following potential explanations.1. ACNH Happy Home Paradise / 2. Apex Legends / 3. Fortnite / 4. Monster Hunter Rise / 5. Diablo II Resurrected / 6. NBA 2K22
![]()
![]()
![]()
![]()
![]()
![]()
On the one hand, this seems to suggest that there still is a bit headroom for microSD and eMMC, but on the other we probably can't expect a leap in performance unless Dane adopts a different technology, such as UFS or NVMe.
- OLED Model's 64GB eMMC is slower than the old 32GB eMMC: Improbable
- OLED Model reads microSD faster than older models: Not sure how to verify this until someone hacks the OLED Model
- Improved sequential transfer speed of EVO Plus: The previous gen was rated "up to 100MB/s", and the latest gen is "up to 130MB/s"
- Better random access speed of class A2: The new EVO Plus is rated A2, while most microSDs are only A1. Class A2 reads 1.67x faster and writes 3x faster than A1. Note that the host device needs to support the A2 standard to take advantage of the speed.
Or it was just an unsubstantiated rumor to begin with.Oh, if it's an old rumor then definitely a lot has changed then.
I feel like (for better or worse) caching/installs kinda have to be the way to go for faster performance. UHS-II is expensive, III basically doesn't exist on the market it seems (might as well be UFS), SD Express is a ways off (and will probably be expensive as well), and I don't see the game cards themselves becoming majorly faster either. Course with heavy usage like caching involves, you'd want big storage or expandable/replaceable memory to reduce the wear.You can get faster eMMC, sure, but you pay more money for it, likewise with UHS-II/UHS-III SD card readers and the cards themselves (there's not all that much to squeeze from UHS-I cards). And for the kind of price they'd pay to squeeze eMMC and microSD for all they're worth, options with far better performance and capacity per dollar start coming into play, including but not limited to UFS.
I just can't see either of the current storage solutions as viable moving forward and I certainly hope they can design a solution to address the game card read speeds, as well.
Razer is making an ARM gaming console. Optimized for cloud gaming. Same gen than 8 gen 1. Maybe 4LPE/5LPP