Hm...for me in Chrome with Google Translate the translation for the middle bullet point is just "finished game consoles"...which is pretty damn smokey. "game console products" could obviously cover more things...things that would likely already be included in points 1 & 3 though. But yeah, even a game console doesn't have to be Drake I guess.
Yes. But more importantly is that Apple isn’t remaining on 5nm for the next few years like we think. They are planning to move on shortly even if Drake launched this year or next year so capacity should be opened up if it hasn’t already.
This could easily just apply for the current switch fwiw. It’s why I don’t consider it as meaning anything tbh.
Oh, certainly. If N3 gets up and running, someone's selling something next year from it (who else would it be other than Apple?).
This isn’t directly related to what you said, and while Digitimes is hit or miss with accuracy, it’s interesting that this is reported:
Note that it could not really refer to a game console but could refer to an accessory that works with the game console
Probably not for Atlan, it’s currently being sampled and it is using “Ampere Next” GPU as per NVidia’s slides. Which is most likely Ampere or Hopper in some capacity I think.
I can see your viewpoint, but then why would they list game consoles AND accessories?Note that it could not really refer to a game console but could refer to an accessory that works with the game console
I’m coping hard here
To sum up thinking from earlier in thread:
Well, there it is. Guess most games that really push the hardware will use DLSS in handheld. Hopefully they pick a good base resolution, going as low as 240p creates significant artifacting.There is good news for those among us who want to see DLSS in portable mode. According to Igors' Lab experiment (I didn't know the source but the tests seem sensible), running Shadow of the Tomb Raider using DLSS actually saves energy per frame rendered:
In this example, the savings amount to 20% power on each frame. That is substantial when we consider what impact on the battery life this can have. It's not gargantuan mind you, but this might have caught Nintendo's attention.
I should say though that all the caveats regarding DLSS in handheld mode still apply: we have less power at disposal compared to docked mode and because of this, DLSS will likely eat up a lot of the time budget per frame. However, if it consistently saves up energy, then it can increase the likeliness of Nintendo trying to make it work.
And of course, the power savings would scale up in docked mode as well.
Thraktor's post here should cover estimates for transistor density/size.
Accessories can apply to pretty much any new product related to the switch family, be it wristband, dock, things like Ring-Fit, etc. they don’t necessarily entail that is new product.
)
Sampling for Atlan actually hasn't started yet. Nvidia plans to sample Atlan in 2023, which is probably when Nvidia has to make the final decision on which process node is used, especially when taking into account the amount of time needed to certify Atlan follows automotive safety standards.
Yes but I’m unsure if this is because they are clocked so high. If they were clocked lower I wonder if this is avoidable and they actually have a similar efficiency to TSMC.
I feel like at the very least, a good number of people here, including myself, believe Nintendo's using Drake, which is a custom variant of Orin, as shown in the leak of confidential Nvidia files by Lapsus$."This is a supercomputer right here in this little motherboard. It can handle input from 12 cameras simultaneously (or a combination of cameras, radar and LiDAR)"
For $10,000, starting in May, developers, universities, and automakers can get their hands on it.
Look at this beast of a card. The first commercial application was for automotive. "a powerful self-driving car computer" "deep learning"
The year was 2015.
If I had told you then that this was going to be the chip in Nintendo's compact handheld console to release in 2 years, would you have believed me then?
The chip was the Tegra X1.
Today you can buy a binned Tegra X1 inside a dev kit for $59! (JETSON NANO 2GB DEVELOPER KIT)
TechPowerUp claims it is a binned Tegra X1, if this forum believes them?
"NVIDIA has disabled some shading units on the Jetson Nano GPU to reach the product's target shader count."
Fast forward to today, and we are discussing if Nintendo can fit an underclocked Orin in a future console.
Some of your reasons against?
-It's too big (still a good point)
-Its too expensive
-It has features that Nintendo does not need, like the deep learning stuff for self-driving cars
-The specs don't match (actually they do)
History shows us that:
-some Tegra products are more versatile then we originally thought
-A chip first announced for use in a product designed for cars with a cost of $10,000, 7 years later can be purchased inside a dev kit with a cost of only $59
Just asking that everyone keep an open mind (within reason of course).
The thing is that the Nvidia hack has already provided the answer to the question of if the next Switch will use the same chip as Orin, and the answer is "no". T239 is clearly labelled as being based on the GA10F die rather than Orin's GA10B."This is a supercomputer right here in this little motherboard. It can handle input from 12 cameras simultaneously (or a combination of cameras, radar and LiDAR)"
For $10,000, starting in May, developers, universities, and automakers can get their hands on it.
Look at this beast of a card. The first commercial application was for automotive. "a powerful self-driving car computer" "deep learning"
The year was 2015.
If I had told you then that this was going to be the chip in Nintendo's compact handheld console to release in 2 years, would you have believed me then?
The chip was the Tegra X1.
Today you can buy a binned Tegra X1 inside a dev kit for $59! (JETSON NANO 2GB DEVELOPER KIT)
TechPowerUp claims it is a binned Tegra X1, if this forum believes them?
"NVIDIA has disabled some shading units on the Jetson Nano GPU to reach the product's target shader count."
Fast forward to today, and we are discussing if Nintendo can fit an underclocked Orin in a future console.
Some of your reasons against?
-It's too big (still a good point)
-Its too expensive
-It has features that Nintendo does not need, like the deep learning stuff for self-driving cars
-The specs don't match (actually they do)
History shows us that:
-some Tegra products are more versatile then we originally thought
-A chip first announced for use in a product designed for cars with a cost of $10,000, 7 years later can be purchased inside a dev kit with a cost of only $59
Just asking that everyone keep an open mind (within reason of course).
TSMC has said their N3 process will have the biggest amount of tape-out in it's first year ever for any TSMC node in history. 2x compared to N5 at launch year. As for which customers will have publicy available designs next year, the ones we know is Apple with the A17 SoC(And maybe M2) and Intel with the integrated GPU Intel Arc Battlemage tile for Intel Core 14th generation Meteor Lake. And maybe Qualcomm with a Nuvia SoC or the Snapdragon 8 Gen 2.
In actuality, based on the breached Information, we know that Orin and Drake are not the same chip at all. ORIN’s Ampere architecture is quite different from the one in Drake. Drake follows the same format found in the Ampere/Lovelace line of GPUs which are gaming ready and are available to PC users. Orin follows the layout of Datacenter Ampere/Lovelace and is meant for high performance in tensor maths."This is a supercomputer right here in this little motherboard. It can handle input from 12 cameras simultaneously (or a combination of cameras, radar and LiDAR)"
For $10,000, starting in May, developers, universities, and automakers can get their hands on it.
Look at this beast of a card. The first commercial application was for automotive. "a powerful self-driving car computer" "deep learning"
The year was 2015.
If I had told you then that this was going to be the chip in Nintendo's compact handheld console to release in 2 years, would you have believed me then?
The chip was the Tegra X1.
Today you can buy a binned Tegra X1 inside a dev kit for $59! (JETSON NANO 2GB DEVELOPER KIT)
TechPowerUp claims it is a binned Tegra X1, if this forum believes them?
"NVIDIA has disabled some shading units on the Jetson Nano GPU to reach the product's target shader count."
Fast forward to today, and we are discussing if Nintendo can fit an underclocked Orin in a future console.
Some of your reasons against?
-It's too big (still a good point)
-Its too expensive
-It has features that Nintendo does not need, like the deep learning stuff for self-driving cars
-The specs don't match (actually they do)
To be fair, the Tegra X1 was designed for mobile first if I’m not mistaken, in NVidia’s venture into the mobile market but that failed and they switch course to the automotive side of things as it was a lucrative business with a lot of potential. They had great visions already previously for Deep Learning before that.
Just asking that everyone keep an open mind (within reason of course).
"Quite different" is a bit of a stretch. There seem to be a few places where GA10F hews closer to desktop Ampere and doesn't have some features of GA10B, namely the double-width tensor cores. Other than that, it is very much a cut down and tweaked Orin, at least on the GPU side.
The reason I say it’s not quite is that it has 1 GPC it seems unlike the 2 in ORIN, and in that 1 GPC it has 12 SMs unlike the 8 SMS in ORIN, and it has 128KB of L1 cache per SM unlike the one in ORiN which has 192KB of L1 cache per GPC. Then there’s the double width Tensor Cores and the higher amount of RT cores/ RT core per SM structure unlike Orin that has 1 RT core per 2 SMs.
Hmmm... maybe.
I dunno. Those things are just different "amounts" of the pieces of the architecture, not differences in the architecture itself. Especially with the cache, increasing or decreasing the size of a component, or the number of components in a cluster, isn't really an architectural change but a configuration change. It still runs Ampere drivers that are set up to be agnostic about such configurations -- a 1 GPC GPU with 6 TPC per GPC is still delivering the same feature set as a 2 GPC GPU with 4 TPC per GPC, and the drivers that back the architecture don't know or care about the difference. There's plenty of flexibility in how the silicon is apportioned without it being a different architecture. And that's exactly what the customization in the original confirmation that "T239 is a customized version of T234" entails.The reason I say it’s not quite is that it has 1 GPC it seems unlike the 2 in ORIN, and in that 1 GPC it has 12 SMs unlike the 8 SMS in ORIN, and it has 128KB of L1 cache per SM unlike the one in ORiN which has 192KB of L1 cache per GPC. Then there’s the double width Tensor Cores and the higher amount of RT cores/ RT core per SM structure unlike Orin that has 1 RT core per 2 SMs.
That doesn’t really sound like a cut down “slightly tweaked” ORIN, that sounds like it’s quite different from how ORIN is.
If tweaked it should follow the general architectural layout of ORIN’s version of Ampere but doesn’t really do that.
I will say though that this is ultimately meaningless in that it’s still part of the Ampere tegra family of SoCs. But the minor semantics of it all, if we are to get technical, doesn’t make it as though it’s 1:1 or close to that. Even using a newer shader model in 8.8 unlike the 8.7 in ORIN or 8.6 in Ampere and it is closer to the Lovelace which is 8.9 in the shader model.
It’s some weird hybrid uArch. But also not in a way.
But the same amount as desktop Ampere.
I can sleep peacefully tonight…maybe.
Yes. 1 RT core per SM for Drake (GA10F) vs 1 RT core per 2 SMs for Orin (GA10B).
Although Nvidia didn't mention on which generation the RT cores on Orin is part of on the Jetson AGX Orin datasheet. So there's a possibility that the RT cores on Orin are of the same generation as the RT cores on Ada GPUs.
Actually, what are the sources for this information? Or even a source for, for instance, the number of RT cores on GA102? Because I'm seeing conflicting details in the source -- places where it says one TTU (tree traversal unit) per TPC and uses that as the RT core count to return to queries, which would be 1 RT core per 2 SMs. But those also seem to apply to all desktop chips with RT cores, whereas the public info on desktop Ampere (e.g. TechPowerUp) states there's 1 RT core per 1 SM.
The Jetson AGX Orin Module Data Sheet, which is now called Jetson Orin Series Module Data Sheet, which no one has access to, unless someone has the Nvidia Developer Program Membership, which I believe @ILikeFeet does. In fact, here's the post from @ILikeFeet.
EDIT: see next post
I'd like to point out that since DLSS 2.3, alot of said artifacting had been resolved. I think the trails in Death Stranding and the after images of Cyberpunk have been fixed?
I'm wondering if this is a case where the marketing concept of an RT core has been used without strict correspondence to the actual hardware, and there's always in fact been 1 TTU per TPC (aka 1 RT core per 2 SMs).
The GA102 whitepaper mentions that the RTX 3080 10 GB Founder's Edition has 68 SMs and 68 2nd Gen RT cores (p 14), and the RTX A6000 and A40 have 84 SMs and 84 2nd Gen RT cores (p 15).
Right, so the full GA102 die is said to have 84 RT cores, matching the SM count.
For the most part I could see dithering relegated to small edges, but Monster Hunter Rise is an RE engine game, so it's entirely dependent on how Capcom devs utilize DLSS.
it does not. there will be a current gen update, and allegedly it'll add RT support, so maybe it'll show up on PC
as far as I heard, PS5's dual sense haptics isn't too different from the HD Rumble in the switch
Sure does feel different to me esp when playing something like returnal
Probably because the haptic motors on the DualSense are physically thicker and larger compared to the HD Rumbles on the Joy-Cons.
