That depends on if Nintendo plans on having the DLSS model*'s motherboard be very similar to the OLED model's motherboard by design, or if Nintendo's going to be designing and manufacturing a much smaller and more compact motherboard that can fit a much larger die for the DLSS model*.
Can't wait for Link's new item, the Industrial Etchant.
this isn't the real Xavier chip?
This is a real die shot of TX1. CPU and GPU are taking a little less than 1/3 of the total chip. We don't have a real Xavier die shot but Nvidia RP 'die shots' from the pascal-turing area were more or less on point for their GPU products. That said I would expect their Xavier/Orin die shots to be slightly different from their PR die shot representations due to their nature of being SoCs with more hardware accelerated parts than GPU only chipsets.
Again! Nvidia does have anything to do with motherboards
That's a 'press die shot'. That said, as I said earlier, most recent press shots from Nvidia have been on point when it comes to real die shots.this isn't the real Xavier chip?
that xavier shot looks way to messy to be a press shot, lol. unless it's when they started to make them accurateThat's a 'press die shot'. That said, as I said earlier, most recent press shots from Nvidia have been on point when it comes to real die shots.
*GA102 real die shot
*GA102 press die shot
Pretty similar but I would be more suspicious over mobile SoCs press representation. They tend to diverge more from their press die shots as they have to deal with more different type of parts than their GPU counterparts.
I was actually editing my comment when you were commenting since I recently woke up when I originally wrote the comment.
Nvidia has heavily pushed for massives SoCs being able to run from credit card board at 10W. It is the sole company that has a 450mm2 chip that runs at 10W that is not a high-end laptop or tablet.
The Xavier shot is a press shot. Just like the Orin one. We have never seen a real die shot of Nvidia SoCs since TX1 that have been pictured twice (first in 2015 and the seconde one in 2017 in order to compare the switch '500 man hours custom' chipset with the original TX1).
Don't usually get that technical with them during discussions.@NateDrake This is kind of random, but have you heard anything from devs about the CPU for the dane? Are they A78c or a78ae? A78c is optimized for mobile gaming and smaller althsn the a78ae in the Orion NX
And I think for now, it's for the best. Right now, I don't think asking for technical details is worth being in hot water with Nintendo.
If not already, it gonna be soon as the Orin Family has been taped out as we have the official specs and they are coming out in Q1 2022.
Outside of process node being used. I still think Samsung's 8N process node is being used to fabricate Dane, especially when taking into account that the max CPU and GPU frequencies of Orin are pretty much identical to the max CPU and GPU frequencies of the Tegra X1.
That's one of the reasons I don't ask. There is little reason to know the CPU and clockspeeds or RAM this far in advance. The info serves no value to the casual audience and the niche details to please a small subsection of those within the community isn't worth the risk or exposure.
Here are the slides for the "Optimizing Deep Neural Networks for NVIDIA DRIVE [SE31472]" session. I think there are many interesting tidbits of information from the slides that I think @Anatole would find interesting and informative, especially the sections about how Orin deals with sparsity, tiling, and chaining.This GTC session introduces Orin SoC's optimization strategies for deep neural networks (DNNs)—sparsity, tiling, and chaining—worth a watch if interested. It also includes the follow slide showing the hardware features that improve DNN optimization.
Edit: typo
To add to this, one thing worth keeping in mind is that the GPU is quite a bit more than the two SMs on TX1, and quite a few of those blocks around and between the SMs are also the GPU, perhaps as much as doubling the size of the GPU. This is pretty important for estimating the size of a small GPU like you'd get on TX1 (or Dane), in that it's not going to scale directly with the number of SMs. There's a certain amount of logic that's going to be per-GPU, a certain amount that's per-GPC, and then the part that's per-SM. For a GPU this small, there's only going to be one GPC, so the first two are effectively static, and it's important to distinguish between those parts and the SM parts. (We've also got per-TPC on the newer architectures, but I'll bundle them into the SMs, assuming the SMs are usually a multiple of two anyway).
This is a real die shot of TX1. CPU and GPU are taking a little less than 1/3 of the total chip. We don't have a real Xavier die shot but Nvidia RP 'die shots' from the pascal-turing area were more or less on point for their GPU products. That said I would expect their Xavier/Orin die shots to be slightly different from their PR die shot representations due to their nature of being SoCs with more hardware accelerated parts than GPU only chipsets.
Problem is what NineTailSage mentioned.
Why are you used unknown numbers of TX1 that has 2 unique SMs and a 7 year old design, vs Xavier, a 8SM Volta GPU which we have exact numbers for. The Volta GPU which includes 64 tensor cores is 89mm² on 12nm. That is the official size of the GPU, even if you assume Ampere has larger GPU logic, you are looking at a process node twice as dense...
Worst case what are we looking at with a 4SM Switch 2?
4SMs at this point is outright unrealistic to me due to how much hoops would need to be jumped through to make it work (Clock speeds, cooler, DLSS wouldn't run well)
also funnily enough it would cost more to do a 4+4 config verus 6 or 8 A78Cs.
where are you getting A55s from? they're no in Orin, so they definitely wouldn't be here
Thraktor
. Then again A78c gives up to 8mb of L3 cache.
www.arm.com
First, it's clear we are discussing Orin NX previously discussed as Orin S, and it's relationship to Dane which uses the same architecture and was developed alongside itself, we know that Nintendo and Nvidia spend R&D into Dane/Orin.
Nano Next was suppose to predate Orin, it was suppose to be on the market right now, only problem is its a 5 watt design and isn't the same architecture as Orin which we know Dane to be.Perhaps Nano Next is Dane/Switch 2, which is a smaller customized version of Orion NX and without the unnecessary A.I. stuff.?
Alright, I'm gonna do some crazy brainstorm b.s. thinking and trying to get a gauge of what Dane could be based off comparing Tegra Xavier at 30 watts vs Xavier NX in 15/ 10 watt modes to Orion NX. Basically what Dane performances hypothetically could be at 15 watts.
Tegra - Wikipedia
en.m.wikipedia.org
If we compare the 30 watt T194 Xavier to Orion at 25 watts.. They both have 8 core CPUs. The Xavier NX 15 watt however has up to 6 core CPUs and 75% of the GPU cores (384 vs 512). Xavier to Xavier NX bus width also lowers from 256 bit to 128 bit.
In this hypothetic scenario I'm gonna a assume that Dane has the same # of GPU cores like Orion NX, but 75% clock speed max. It's essentially also going to be the same clock speed or close to Switch's (768 Mhz) docked mode, which would give us around 1.5 TFLOPs.
And then lower the CPU count from 8 A78AE cores to A78c 6 core CPU. A78C also happens to be smaller then the A78AE as well. Also, I'm gonna say the 128 bit bus bandwidth stays.
So somehow I can see Dane/Switch 2 be a 1.5-1.6 TFLOPs GPU, 6 A78C CPUs (1.5 GHz max?),
128 bit bus width with 8-12 GB lpddr5 RAM using 128 bus width to achieve 102 GB/s bandwidth at 15 watts power on a 6-7 inch 720p screen.
And so somehow all of this is on a chip as similar in size as TX1. I don't even accounted included DLSS or the larger memory cache. Also, no comment on RT cores.
When you think about it, it's really impressive thinking we are capable of matching GCN PS4 GPU perfomance on an 8nm node. And it's pretty cool when we see that ampere at 8nm is very comparable to 7nm rdna2 in performance. And guess what.. Steam Deck runs up to 1.6 TFLOPS GPU and 128 bus width lpddr5, all at 15 watts power like OG switch.
And man, Switch 2 with DLSS and the extra memory cache would be insane if they could fit it all in.
Fair enough. Since we are ruling out that the Dane is based off the Orion NX because its too big (like Xavier NX), I think the Orion S can still exist now, despite it not showing up this week's presentation. We got thrown off thinking the Orion S just consolidated and merged into the NX.. But Maybe Nvidia really left this out on purpose because they didn't want us putting 2+2 together, until the time was right.. They still have more to reveal by the end of the year or q1 2022, don't they
? Nintendo would be the ones officially announcing the switch 2 of course.
Orin NX chip size is unknown, don't rule it out. They aren't making Orin NX chips, they are making Orin chips around the same size as Xavier, and binning some as Orin NX. Nvidia is not making Orin NX dies, They would be at best 60% the size, and Dane would lose another ~20% removing all the AI accelerator engines.Fair enough. Since we are ruling out that the Dane will based off the NX because its too big (like Xavier NX), I think the Orion S can still exist now, despite it not showing up this week's presentation. We got thrown off thinking the Orion S just consolidated and merged into the NX.. But Maybe Nvidia really left this out on purpose because they didn't want us putting 2+2 together, until the time was right.. They still have more to reveal by the end of the year or q1 2022, don't they
I think we are all in the ball park on what Dane's power could be. We've discussed this before.. The 1.6 TFLOPs and 6-8 core A78 CPU using 128 bus width of lpddr5 doesn't seem impossible. AMD and Steck are already matching this on 7nm on a form factor slightly bigger than switch, and with the same 15 watts usage.
You don’t need 4 to do that task, it is supposed to be using the same OS as the switch is correct? Which is Horizon OS. The clocks for a single A55 core, which is significantly smaller than an A78 might I add, would only need to match what the OG switch needed in terms of performance, 1.3GHz is enough to do the same task as 1A57 at 1GHz while consuming a fraction of the power draw and taking up a fraction of the space for it.
Why did you remove 2 CPU coresPerhaps Nano Next is Dane/Switch 2, which is a smaller customized version of Orion NX and without the unnecessary A.I. stuff.?
Alright, I'm gonna do some crazy brainstorm b.s. thinking and trying to get a gauge of what Dane could be based off comparing Tegra Xavier at 30 watts vs Xavier NX in 15/ 10 watt modes to Orion NX. Basically what Dane performances hypothetically could be at 15 watts.
Tegra - Wikipedia
If we compare the 30 watt T194 Xavier to Orion at 25 watts.. They both have 8 core CPUs. The Xavier NX 15 watt however has up to 6 core CPUs and 75% of the GPU cores (384 vs 512). Xavier to Xavier NX bus width also lowers from 256 bit to 128 bit.
In this hypothetic scenario I'm gonna a assume that Dane has the same # of GPU cores like Orion NX (1024 cuda cores), but 75% clock speed max. It's essentially also going to be the same clock speed or close to Switch's (768 Mhz) docked mode, which would give us around 1.5 TFLOPs.
And then lower the CPU count from 8 A78AE cores to A78c 6 core CPU. A78C also happens to be smaller then the A78AE as well. Also, I'm gonna say the 128 bit bus bandwidth stays.
So somehow I can see Dane/Switch 2 be a 1.5-1.6 TFLOPs GPU, 6 A78C CPUs (1.5 GHz max?),
128 bit bus width with 8-12 GB lpddr5 RAM using 128 bus width to achieve 102 GB/s bandwidth at 15 watts power on a 6-7 inch 720p screen.
Maybe this is possible at 15 watts? 25% less CPU cores and they are clocked 25% lower, along with 25% lower GPU speeds and take out the camera and other unnecessary stuff from Orion NX. And so somehow all of this is on a chip as similar in size as TX1. I don't even accounted included DLSS or the larger memory cache. Also, no comment on RT cores.
When you think about it, it's really impressive thinking we are capable of matching GCN PS4 GPU perfomance on an 8nm node. And it's pretty cool when we see that ampere at 8nm is very comparable to 7nm rdna2 in performance. And guess what.. Steam Deck runs up to 1.6 TFLOPS GPU and 128 bus width lpddr5, all at 15 watts power like OG switch.
And man, Switch 2 with DLSS and the extra memory cache would be insane if they could fit it all in. Only thing is if they use A78c, they might not necessarily get the 3MB L2 + 6MB L3 cache upgrade
Cortex-A78C
Providing market-specific solutions with advanced security features and large big-core configurations
I was basically making some sort of an educated guess on predicting Dane's specs (10-15 watts) comparing it to the 30 watt Volta and noticed the 10-15 watt Volta versions had 6 CPU cores instead of 8. Not to mention slower clock speeds as well. Also retained 75% of the cuda cores, and at slower clock speeds and a 128 bus width 8 GB RAM (instead of 256 and more RAM) in an effort to meet the power draw requirements of 10-15 watts.You don’t need 4 to do that task, it is supposed to be using the same OS as the switch is correct? Which is Horizon OS. The clocks for a single A55 core, which is significantly smaller than an A78 might I add, would only need to match what the OG switch needed in terms of performance, 1.3GHz is enough to do the same task as 1A57 at 1GHz while consuming a fraction of the power draw and taking up a fraction of the space for it.
This, of course, is a completely custom setup and won’t happen really, having a single A55 core. 2 is possible imo if they want a set of cores dedicated just to OS stuff. If there are 2 A55s they can be clocked lower and it would consume even less power.
This in turn prevents the issue of say 6A78 cores where 1 is used for the OS and only 5 are available to games. The set up with little + big is to free up as much of the A78 to be clocked a lot higher and leave the OS to the little cores and taking up even less space that can be allocated to other things like more cache perhaps. While you can use an A78 in a 6 core setup, you leave a noticeable amount of performance that can be allocated to a game directly.
In theory, you can get a much better battery life and an even more performant device in its lowest case and it’s highest case scenarios, while not consuming more die space and fitting within a smaller budget and still keeping 6-8SMs intact.
On another note, since Dane will be based off an architecture that has RT cores, those RT cores can be used for Audio in theory if they aren’t using it for docked or portable mode.
Why did you remove 2 CPU cores
I'm discussing 4xA78+4xA55 cores vs 6xA78C cores. A55 cores is more of a want from how people would build the next Switch, than we are seeing from Nvidia, who hasn't used A53 or A55 cores in any design since the Tegra X1's days. Please take into account that A78C cores are designed for Dedicated gaming devices, and that developers would expect 8 high performance cores, so they would expect 8xA78C cores, and that is what we are seeing in Orin, the whole big.little thing is on us, not any rumors or found evidence, no hint that Nvidia or Nintendo are going that route. Also because you can clock a core separately from the others, you could run 7xA78C cores at 1.6GHz and 1xA78C core at 450MHz, it takes up the same size as 4xA55 cores, and the same performance as those cores at 1.35GHz, so power consumption should not be as much as an issue. This allows 8 cores on 1 cluster with the same memory cache as 8xA78AE cores, which would be nice because they would share the same code, similar register/cache space.You don’t need 4 to do that task, it is supposed to be using the same OS as the switch is correct? Which is Horizon OS. The clocks for a single A55 core, which is significantly smaller than an A78 might I add, would only need to match what the OG switch needed in terms of performance, 1.3GHz is enough to do the same task as 1A57 at 1GHz while consuming a fraction of the power draw and taking up a fraction of the space for it.
This, of course, is a completely custom setup and won’t happen really, having a single A55 core. 2 is possible imo if they want a set of cores dedicated just to OS stuff. If there are 2 A55s they can be clocked lower and it would consume even less power.
This in turn prevents the issue of say 6A78 cores where 1 is used for the OS and only 5 are available to games. The set up with little + big is to free up as much of the A78 to be clocked a lot higher and leave the OS to the little cores and taking up even less space that can be allocated to other things like more cache perhaps. While you can use an A78 in a 6 core setup, you leave a noticeable amount of performance that can be allocated to a game directly.
In theory, you can get a much better battery life and an even more performant device in its lowest case and it’s highest case scenarios, while not consuming more die space and fitting within a smaller budget and still keeping 6-8SMs intact.
On another note, since Dane will be based off an architecture that has RT cores, those RT cores can be used for Audio in theory if they aren’t using it for docked or portable mode.
Why did you remove 2 CPU cores
Well, that's a bit of an unknowable variable. It's right to assume it will be as lightweight as possible like Horizon, but whether or not it's the exact same? Who knows.
Thraktor's post was comparing TX1's unknown numbers to Orin's unknown numbers... it didn't make much sense to be honest. He should have used Xavier's known numbers, we know what Xavier's 8SM GPU takes in terms of size, it is 89mm^2 officially, for the GPU in Xavier, that is at 25M transistors per mm. Dane would be ~50M transistors per mm, and suggesting a ~50% increase in transistor count gets you to 66mm^2, if the GPU is ~40% of the die, that gives you a ~160mm^2 die that is worst case IMO, and is a similar size to Wii U's entire GPU die, so it isn't unreasonable. Realistically, Xavier has 64 tensor cores and Orin NX has 32, Xavier has fp64 hardware, Orin NX does not... I'd suggest only a 25% transistor increase here, based on how big Volta was, with these numbers and the GPU being 40% gives you a die area under 140mm^2. Considering the higher price point, I think it makes sense...
That article is giving off multiple red flags.
This is purely hypothetical, but I don't think half of docked clock is a hard limit for Dane. Nintendo used the 460 MHz clock profile on Erista out of all things for their high profile games and that more than half of the docked clock profile (768 Mhz). That said we may never know about Dane's clock profiles till it's released.
Or both I'd guess. At least for Erista, Nintendo opened up the options for 3rd party developers to use either the 372 or 460 Mhz profiles in handheld mode (and possibly the CPU OC mode but Idk if any one used them so far), so I hope that such freedom is maintained for Dane.
I wasn't suggesting the info was legit, but the idea that the portable GPU would have a higher clock than expected, because it doesn't use DLSS, could be a trade off Nintendo explores... for instance, currently Switch uses a 768MHz docked performance and a 460MHz portable performance (60% of docked performance)... If Gen1 Switch could use DLSS, then it would make sense to have docked and portable clocks even closer together (say 80%) this way you can render in 720p and output much higher resolutions when docked, but stay at 720p when portable... Which Dane still might be 720p.
It doesn’t really make sense though, the Volta you were comparing to is the non-NX which is 30w and has all 8 cores, the 15w Volta has 1/4th of the cores disabled and that is the NX.
You can clock a core separate from others, but the cost of doing so is a more complex design to do such a thing. You don’t simply clock it lower, it’s designed to be clocked lower by having a separate railing for its required voltage.
Yes, 450MHz was one example, it could be 500MHz for instance, but doing this to a single core is less complex than a separate cluster with its own rail system. Technically you can do it for all 8 cores, but since you'd only do this for the OS core, I think doing it once makes the most sense IMO.It doesn’t really make sense though, the Volta you were comparing to is the non-NX which is 30w and has all 8 cores, the 15w Volta has 1/4th of the cores disabled and that is the NX.
You were comparing the non-NX to the NX from Volta to the Orin and basing the conclusion on that, which is inconsistent to make.
You basically removed more cores for no reason
It would have 8Cores, not 6 cores.
The other things were lowered the right way, the CPU count was unnecessarily lowered.
I doubt it would be a heavy OS, past Nintendo system with the heaviest OS was the Wii U, other systems had a non-existent OS, Switch was made with being as lean as possible on purpose, I doubt they will try to make the OS heavy like the other systems.
You can clock a core separate from others, but the cost of doing so is a more complex design to do such a thing. You don’t simply clock it lower, it’s designed to be clocked lower by having a separate railing for its required voltage.
Especially for something that is so far in clicks like 450MHz and 1.6GHz.
No, Orin NX is Orin S, look at the picture above. Dane is what Nintendo will use, and that isn't Orin NX, it removes those AI components, but it still will have a 10-15 watt TDP and use the same architecture, core counts don't really matter when talking about Dane, it will be the same performance per watt as Orin NX outside of the automated engine stuff.
