May I note that a lot of Orin's size is using A78AE's and DLAs, and other extraneous parts due to its nature as an AI/Automotive chip?
I keep seeing this claim. Is there a source to the size difference?
There's no indication on how much bigger it makes the cpu cluster
Not a single source has said they were unhappy with 8nm.
I completely forgot we had a die shot of Orin.If Orin is indeed 460mm2, that would mean 1 A78AE is roughly 45mm2
So 1 A78AE is around 1% of the die size of Orin. and the whole block of 12 A78AEs on 8nm is around 155-156mm2.
That is the Dimensity 1200, a chip that while not having an official die-size revealed, looks to be at most the size of a Cheez-it which is around 24mm2.
The Dimensity has 4 A78s, and 4 A55s, and while the A55s, and GPU assembly make the chip overall smaller to fit into that "around 20-30mm2" range, the important thing is the 4 A78s.
While we have no die shot, we can extrapolate a bit of the size difference here.
Samsung 8nm that Orin is fabricated on is a tad less Dense than TSMC 7nm, the Dimensity 1200 is fabricated on TSMC 6nm which is 18% denser than 7nm.
So let's highball it and say TSMC 6nm is 30% Denser than Samsung 8nm.
That would mean a Single A78AE on TSMC 6nm, the same fabrication that has a SoC that has 4 A78s with 4 A55s and a GPU of its own.
Would eat up the majority, or all the area of the Dimensity 1200 with the "6nm A78AE" being over 30mm2
You'd need to say that TSMC 6nm is 85-95% denser than Samsung 8nm in order to fit 4 A78s into that if the A78s were the same size as the A78AEs. (At least assuming the GPU takes up the majority of space in the Dimensity like it does in some other SoCs)
I have no idea how you reached that estimated area for a single A78AE core. My rough estimate working from the whole image being 21mmx21mm is 5.75mm² for a single core.If Orin is indeed 460mm2, that would mean 1 A78AE is roughly 45mm2
So 1 A78AE is around 1% of the die size of Orin. and the whole block of 12 A78AEs on 8nm is around 155-156mm2.
That is the Dimensity 1200, a chip that while not having an official die-size revealed, looks to be at most the size of a Cheez-it which is around 24mm2.
The Dimensity has 4 A78s, and 4 A55s, and while the A55s, and GPU assembly make the chip overall smaller to fit into that "around 20-30mm2" range, the important thing is the 4 A78s.
While we have no die shot, we can extrapolate a bit of the size difference here.
Samsung 8nm that Orin is fabricated on is a tad less Dense than TSMC 7nm, the Dimensity 1200 is fabricated on TSMC 6nm which is 18% denser than 7nm.
So let's highball it and say TSMC 6nm is 30% Denser than Samsung 8nm.
That would mean a Single A78AE on TSMC 6nm, the same fabrication that has a SoC that has 4 A78s with 4 A55s and a GPU of its own.
Would eat up the majority, or all the area of the Dimensity 1200 with the "6nm A78AE" being over 30mm2
You'd need to say that TSMC 6nm is 85-95% denser than Samsung 8nm in order to fit 4 A78s into that if the A78s were the same size as the A78AEs. (At least assuming the GPU takes up the majority of space in the Dimensity like it does in some other SoCs)
given that RT seems to "work" (however much that means), we're probably looking towards a higher SM count
*MX550(Unless Nvidia brands it as RTX 3030). It's probably going to be half/third of the GA107 die. Considering the MX line is their answer to whatever the strongest iGPU at the time is, I think they will aim at close to GTX 1650 performance at 28.5W.
Isn't 8 the current at the moment? With 7 being bleeding edge and 5 being the next big thing? Surely we don't expect the next hardware to be bleeding edge.
so will it have RT cores and tensor cores? given there's no non-tensor Ampere, it sounds likely. maybe the MX550 is just a laptop version of Orin NX
no, 8nm is old. it's just a better 10nm. 7nm is passe now (PS5 and Series use it, Ryzen use it, RDNA1 and 2 use it, etc). 5nm is already in products and will see more products early next year. late 2022, we might see 3nm products
So I was bored and was playing around with a new app I got on my iPad...
Well for one Orin is actually a square based on the Orin NXSo I was bored and was playing around with a new app I got on my iPad...
Basically just scaled the die shot to come out to ~460mm2, then got the rest from there, about ~4mm2 for the A78AEs (so of course ~48mm2 for the x12 cluster of them). Course it's all an in exact science based off a jpeg, like depending how it was cropped and how they're counting the size would mess with all those measurements.
Nothing really notifying much change to my knowledge no.
I don't think there are going to be any late 2022 products using 3 nm** chips, considering that TSMC said that 3 nm** wafers won't be shipped until Q1 2023, and Samsung's 3GAE process node won't start mass manufacturing until EoY (End of Year) 2022, which seems to only be available for Samsung, and Samsung's 3GAP process node won't start mass manufacturing until EoY 2023.
I think it's the daughterboard, not the motherboard, that's heavily modified. Going by IFixIt's teardown of the OLED model and the Nintendo Switch, the OLED model's motherboard still looks quite similar to the Nintendo Switch's motherboard.The interesting look is at how much Nintendo has already modified the original Switch motherboard to accommodate the OLED revamp...
And, as they mentioned, the thermal management shrank a fair bit. Even to the naked eye, you can see smaller fans and heat pipes.I think it's the daughterboard, not the motherboard, that's heavily modified. Going by IFixIt's teardown of the OLED model and the Nintendo Switch, the OLED model's motherboard still looks quite similar to the Nintendo Switch's motherboard.
OLED model:
Nintendo Switch:
To be fair, the Nintendo Switch's fan and copper heat pipe was complete overkill for the Tegra X1+, considering how much more power efficient the Tegra X1+ is compared to the Tegra X1.
it's very unlikely Dane is gonna come close to the 1650S
famiboards.com
I was never saying that it would come close to the 1650S Natively XD.
It was slight overkill even for Erista, if we take iFixit at their word. Which is why I was thinking 15W TDP is the moderate level of power consumption Nintendo might target for a new chip, but couldn't go much higher than 20W at the absolute maximum.
I just realized that annotated image is based off this off angle render/press image from however long back, late 2019 from a quick search.Well for one Orin is actually a square based on the Orin NX
70mm by 45mm for the whole thing.
Maybe try plugging in that into your app to figure out how big the main Orin die is?
If so then we could actually figure out how big/small the full Orin Die as Orin NX is a binned Orin Die (so Orin proper is the same size).
Either way, Orin NX is quite tiny for what it's packing, and even if the A78AE's are only marginanlly bigger (they are bigger as they have some sort of extra logic for AI/Auto), cutting off 4 Cores slashes it down to 24mm2 for the A78AE's if your Orin-Dieshot number is accurate (Will have to check it against Orin NX's proper dimensions).
So 8 A78C's would likely be 20-22mm2.
And then cut the GPU die size in half (how big is the GPU there?)
With Orin NX's image and exact measurements, Orin's Dieshot (As Orin NX has the same die), and that app, we could actually figure out how much space 8SMs, and 8 A78C's would take up.
Seriously though, Orin NX is small enough to be Dane literally board size-wise if they tweaked the internals more than a plug-in solution as Orin NX has most of what is needed for a system.
It gives me bad memories of the old threads where we did the same work with Xavier.So I was bored and was playing around with a new app I got on my iPad...
Basically just scaled the die shot to come out to ~460mm2, then got the rest from there, about ~4mm2 for the A78AEs (so of course ~48mm2 for the x12 cluster of them). Course it's all an in exact science based off a jpeg, like depending how it was cropped and how they're counting the size would mess with all those measurements.
Yeah, we can't trust at all that's the real Orin in 2019 nor that it didn't change significantly in those 2 years.
That's wrong. Most modern SoCs have a GPU+CPU ratio that is 1/3 of the overall chip. That's true for TX1, Xavier, E990 and Apple SoCs (those that don't have an integrated 5G modem like 7 nm Kirin SoCs where the ratio is closer to 1/4). By your calculations, Dane would be 264 mm2 which is more than 2*TX1.
They can use a very wide 400mm2 chip. It just need to run under 10W with the actual OLED battery/cooling solution.
To be frank we don't even know if it'll have the same form factor as the Switch (original or OLED) so yeah I don't think they're beholden to any one die size.
Maybe I'll wait for the Dane version in 4 years lol.
Rather than GPU+CPU, don't you mean CUDA Cores + CPU? Because from a quick check on the TX1, the 3 rectangles containing CUDA Cores and Arm cores is roughly 31.5% and that's without any margin. I honestly can't identify the remaining components of the GPU there, so I can't say how much is the CPU+GPU percent is, but if we're only looking at CUDA Cores in the ORIN die shot, they're roughly 8x169x161 ~= 26mm2.
If we extrapolate from that 1/3 rule that would be 155~160mm2. Though does that rule really stands for smaller nodes? Don't phones SoCs have anything else integrated in the SoC which a gaming console doesn't needs to?
Whoa, that's completely different from that press image. In that case, that Orin press image should be pure bullshit as well.
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.
This makes me think of Wooded Temple in Spirit Tracks, where you have to use the whirlwind to puff away the toxic clouds that are covering the place.
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.
