-
Hey Guest. Check out your NeoGAF Wrapped 2025 results here!
A Nintendo Switch has been taken apart
- Thread starter Dystify
- Start date
Elfotografoalocado
Banned
They better have the goods man.
They identified the parts of the Apple A10, maybe it's different to TX1 and they took longer to identify the parts?
They identified the parts of the Apple A10, maybe it's different to TX1 and they took longer to identify the parts?
LordOfChaos
Member
They're giving it for free, again, when they charge a buttload for die scans and analysis, just because we were interested. Lets not give them snark here
Elfotografoalocado
Banned
Well, the blog is part of their commercial strategy for sure. It gives them publicity and is a sample for the work they can perform as a comission.
Xdrive05
Member
Does anyone want to make final predictions about what's under the lid?
I'll be conservative and predict it's a standard 20nm X1 with A57s and the A53 cores have been zapped (disabled), with no extra cache at all. Nintendo is relying on Nvidia color compression and TBR to make up the gains they would usually achieve from cache and embedded memory modifications/additions.
Either I'm right, or I'm very pleasantly wrong.
I'll be conservative and predict it's a standard 20nm X1 with A57s and the A53 cores have been zapped (disabled), with no extra cache at all. Nintendo is relying on Nvidia color compression and TBR to make up the gains they would usually achieve from cache and embedded memory modifications/additions.
Either I'm right, or I'm very pleasantly wrong.
Thanks for posting this, it's really helpful to have an original TX1 die photo to compare against for Switch, so it's great to see Fritzchens Fritz has done one just in time.
For reference, this is from the original Shield TV, not the new 2017 version (they have different die markings, so there may be some small changes to the TX1 in the newer Shield TV).
I started doing up a diagram of where everything is on the die photo, to use as a point of reference against a Switch die shot... but then I got distracted and started playing Zelda instead. So here you go:
LordOfChaos
Member
Shield TV TX1 does include A53s, they're just disabled (no battery, larger space for cooling, they just chose responsiveness over slight power savings)
Elfotografoalocado
Banned
My optimistic guess is:
16nm
4xA72s
2xPascal SMMs
Modified uncore for cache coherence (Standard ARM uncore, licensed)
16nm
4xA72s
2xPascal SMMs
Modified uncore for cache coherence (Standard ARM uncore, licensed)
anchorman128
Member
They're giving it for free, again, when they charge a buttload for die scans and analysis, just because we were interested. Lets not give them snark here
100% agree, although I am sad that we are only getting a top metal die shot. Won't be able to discern 16nm vs 20nm with that alone. However, we may be able to make a educated guess based off what the rest of the die reveals.
But hey, maybe they are getting us a full on scan with an electron microscope? One can only hope.
There was someone who said he'd done it http://www.neogaf.com/forum/showpost.php?p=231595958&postcount=485
It was from this thread Joycon connection issue may be hardware related
Original video of this modification of the left Joy-Con can be seen at https://youtu.be/ZnMnke6lF0c
LordOfChaos
Member
Nice work. The other blocks would be this stuff, just out of Nvidias heavily stylized order:
There's at least two blocks Nintendo may not need or not need such a large version of, but at the estimated die size it would be odd coincidence if they removed a lot and added some stuff and ended up the same size, from what we can tell (still waiting on calipers). If the block capable of 4K decode for instance is still in there, that's certainly of interest (though the USB 3.0 datarate to the dock may make that moot?)
Elfotografoalocado
Banned
All of that is inside too, right? The thing is supposed to record gameplay after an update so presumably that'd be based on Shadowplay tech?
Why do you guys think BOTW Switch in docked mode was only 900p with significant sub 30 fps drops?
Best the Switch could do? Different architecture from the Wii U made it difficult to do better and/or Nintendo didn't want to spend too much time on the port? Nintendo is still learning the Switch architecture and there will be better graphical games in the future?
Best the Switch could do? Different architecture from the Wii U made it difficult to do better and/or Nintendo didn't want to spend too much time on the port? Nintendo is still learning the Switch architecture and there will be better graphical games in the future?
Look at this die analysis, we can spot the same formation of a57s although it seems in a less wider fab process, but there is no little a53s "cross" cluster in Shield TV die.
Some time ago I found a page stating that a53 were ditched from later X1s but I just can't find it now :/
PD.- Needed to rehost properly the img sorry.
My final bet is that Switch SoC is 95% to 100%
the same as in Tegra Shield 2017 (meaning It may be the same chip), but I'm open to the possibility of not being on 20nm process, I also think there are some slight differences between Shield TV 2017 SoC and the old one.
one the one hand, that seems like a lot of space wasted on boring crap - you could squeeze more SMs in there.
On the other hand, once they die shrink, hopefully those boring bits will get smaller, so if they keep a similar die size then then can put more SMs in future versions.
On the other hand, once they die shrink, hopefully those boring bits will get smaller, so if they keep a similar die size then then can put more SMs in future versions.
LordOfChaos
Member
Look at this die analysis, we can spot the same formation of a57s although it seems in a less wider fab process, but there is no little a53s "cross" cluster in Shield TV die.
Some time ago I found a page stating that a53 were ditched from later X1s but I just can't find it now :/
Can't see the image, but afaik the new and old Shield TV used the exact same chip, no difference in revision ID or anything, just with different cooling setups that made the new one a bit better.
Only references I can find to A53 removed + Tegra X1 are speculation on the Switch.
Would be interesting if someone could confirm the A53s were removed from a version of the TX1 though of course, and I'd be interested in what the new die size would be.
Gotdatmoney
Member
The pics we got from before, dont they basically disprove 28nm process?
Schnozberry
Member
I'd contact Nintendo about a replacement before I voided the warranty. I have two sets of joycons and they work from 20 ft away behind my back.
Elfotografoalocado
Banned
The hypothesis I made was, the game is very bandwidth-heavy because of its physics and alpha effects, and since it comes from a system based around having the whole render targets in its 32MB eDRAM, whereas Switch uses tiles in order to have a smaller, cache based, memory pool. Since Zelda was originally designed for Wii U not all of the graphics can be tiled and the game suffers in bandwidth intensive areas and when streaming assets.
I know, having low hopes
Shhh
I rather have low hopes and be surprised than disappointed
Xdrive05
Member
That is sound speculation. I'd be curious to see what a BotW caliber game could be like if actually made for the Switch from conception. The Switch is so much more capable than what BotW shows. The architectures are just too different for the port to be playing to Switch's strengths, or more appropriately, playing away from it's bandwidth limitations.
Port from a completely different architecture made mostly on non final hardware and rushed for launch.
Nintendo: The internals is a Custom tegra optimised for performance and power efficiency.
Devs: Great! What are these customisations?
Nintendo: We took a TX1, enlarged the fabrication node, and put in a much less efficient and weaker cpu.
Devs: Wtf.
Kimishima: excellent!
Nintendo: Uuuh we made Switch more expensive and have shorter battery life!
Kimishima: But Zelda will run perfectly right?
Nintendo: Nope...
Kimishima: DO'H
LordOfChaos
Member
Phoronix TX1 benchmarks may be of interest in the meantime...If you get all 4 cores spinning it's not worlds apart from an i3 in some cases, wedged between that and bay trail.
http://www.phoronix.com/scan.php?page=article&item=nvidia-tegra-x1&num=2
http://www.phoronix.com/scan.php?page=article&item=nvidia-tegra-x1&num=2
Elfotografoalocado
Banned
I mean, the transistor count of the chip would barely decrease, and wouldn't account for the huge die size that 28nm would entail.
Also, the redesign would be as expensive as making it more powerful, and the 28nm chips would not be significantly cheaper than 16nm ones.
It makes no sense, engineering wise, commercially wise, or in any way.
Also, the redesign would be as expensive as making it more powerful, and the 28nm chips would not be significantly cheaper than 16nm ones.
It makes no sense, engineering wise, commercially wise, or in any way.
LordOfChaos
Member
Pretty sure 28nm was just a joke at Nintendoings
20nm seems reasonably sure, with a mild hope for 16nm FF. Both are near the same transistor density (aint non-standardized node measuring and advertising grand?)
I'm not sure about bandwidth being the whole problem considering Switch in handheld mode can run Zelda better than WiiU. Perhaps its a small part of the issue though. The extra 20% bandwidth going from handheld to docked not quite matching the increased bandwidth requirements of the increase from 720p to 900p. Causing a small framerate drop (say 30fps down to 27/28fps) and double buffering causing that to stutter to 20fps.
Elfotografoalocado
Banned
The way I see it, of the devices in the same range as Switch released in the last 18 months, the 100% of them are in 16/14 nm and the 0% of them have A57 CPUs.
So, why would Nintendo use them?
Nvidia had 250 of its 10000 employees working exclusively on this thing, I'm sure that's enough for designing an updated chip compared to the 2 year old TX1.
Edit: about Zelda, that's the thing. Switch has 20GB/s of main memory bandwidth available in portable mode, so in that mode it has no problem because of the resolution difference.
So, why would Nintendo use them?
Nvidia had 250 of its 10000 employees working exclusively on this thing, I'm sure that's enough for designing an updated chip compared to the 2 year old TX1.
Edit: about Zelda, that's the thing. Switch has 20GB/s of main memory bandwidth available in portable mode, so in that mode it has no problem because of the resolution difference.
A resolution increase wouldn't increase main memory bandwidth linearly though. Framebuffer access in main memory (especially for a TBR) will only be a fraction of total main memory usage. So the 56% resolution increase wouldn't come close to requiring 56% more bandwidth. It may need more than the 20% increase docked mode gives Switch though. So like I suggested maybe a small framerate dip caused by bandwidth constraints with double buffering then causing the major drop to 20fps?
Schnozberry
Member
It might have something to do with quickly porting an engine that was designed for Wii U somewhat late in the development cycle.
LordOfChaos
Member
Because no 16nm Tegra is available as of yet. What other companies have done with their SoCs doesn't mean Nvidia has come along. And A72 and Pascal aren't taped out on 20nm.
250 people working two years could certainly do /something/, but that's not a massive amount of man-hours to radically redesign a tegra in the chipmaking world.
The die size is also suspiciously very similar to the TX1, so occam's razor here,
1) They reconjiggered a bunch of stuff, moved to a new node, and coincidentally landed on the same die area
or
2) Same node, same basic layout of CPU/GPU, minor changes here and there where Nintendo found interest (i.e memory chain)
We'll see hopefully shortly, but I think expecting a radically different chip than TX1 is setting yourself up for a letdown.
No doubt, we're just wondering about the specific cause in this port.
R
Rodin
Unconfirmed Member
Eh, i've been saying this since the DF analysis dropped, Tom said that too, nobody's listening.
There also seem to be some issue with the streaming (sometimes when it starts dropping if i stop and start moving again after 1-2 seconds it doesn't dip anymore) and possibly a memory leak somewhere, which would potentially explain the "game stopped working due to an error" message that some of us got (happened 10 times to me before i could actually start it).
This too, but it never drops in 720p so better bandwidth management in 900p should bring performances to undocked levels. GPU certainly can't be an issue.
Gotdatmoney
Member
I know, having low hopes
Shhh
I just abandoned hopes long ago. I just wanna know cause this whole ride has been a months long wild ride haha.
If it's just a bandwidth problem shouldnt the WiiU version perform similar to the the handheld switch performance? The Switch docked performance is similar to the WiiU. I don't think there is much we can tell from what we currently know besides double buffering causing the big ass drops.
I honestly dunno with they used this low ass memory bandwidth ram regardless of whether its the issue with Zelda or not.
Elfotografoalocado
Banned
The thing is, if the change is from A57 cores to A72 cores, same memory controller, different uncore, 20 to 16nm, the difference in size would be negligible because the components would have very similar transistor counts and the density of both processes is the same.
Furthermore, what pushes Nvidia to invest in designing new Tegra chips is the possibility of revenue. A big console contract is exactly that. The fact is, Nvidia 10000 employees for everything they do, and 250 of those employees spent two whole years on this chip for a breakthrough contract that will decide the fate of future semi-custom chips for Nvidia.
How would possible clients think of the company if, after paying millions of dollars for the design, they gave them an already existing 2 year old chip on a failed, expensive process node?
Parker is a 16nm tegra. It's made for cars but still. Would a node shrink really count as a radical redesign?
I agree with you from all the evidence it's most likely 20.
LordOfChaos
Member
Afaik they detailed the layout, but Parker isn't in any shipping products yet, or has that changed?
Wasn't calling the node shrink itself a radical redesign, but rather that node difference combined with changing all the functional units on a TX1 and coincidentally landing on a similar die size would be a big coincidence, rather than the seemingly more likely scenario of it being on the same 20nm node, same functional units, a tweak here and there.
The 250 people x 2 years also includes the bespoke NVN API, the OS integration, some sort of physics library tailored for Tegra they mentioned, etc. Doesn't seem to leave much room for changing things a whole lot.
Elfotografoalocado
Banned
The point I want to make is, before PS4 and Xbox One were released, there were no octacore Jaguar, high GPU power APUs on the market with GPU featuresets ahead of those in regular GCN GPUs of the time and with full cache coherence. Sony and Microsoft paid millions of dollars for those chips to happen, and now Nintendo has done the same.
Otherwise Nvidia has no reason to spend the engineering hours in such a chip, so why would they release it before the target device for a chip of those characteristics?
Otherwise Nvidia has no reason to spend the engineering hours in such a chip, so why would they release it before the target device for a chip of those characteristics?
According to this website, all Teslas manufactured after October 2016 features PX2 Parker.
https://electrek.co/2016/10/21/all-...-new-drive-px-2-ai-platform-for-self-driving/
tkscz
Member
Why is this so hard for people to get? The Wii U has an IBM Power based CPU, and a VLIW5 based AMD GPU (Close to the HD Radeon 4000 series) and Zelda barely ran on it, and it was ported over in less than a year on non-final Nvidia Maxwell 2.0/ARM57x hardware and forced to a higher resolution and didn't even run perfectly on previous hardware. And bringing up Mario Kart 8 Deluxe running at 1080p 60fps means crap as it's nowhere near as large.
Woah so I'm not the only one who noticed this.
But it actually rang a bell thanks to the WiiU version (I double dipped), where I cpoukd actually correlate that to the disk noise, whereas this was obviously not possible with the cart version.
Elfotografoalocado
Banned
That's what TBR is for, with Vulkan you can even do full deferred rendering without touching main memory just like you do in Wii U. If the thing is cache coherent, that also reduces bandwidth needs by eliminatig copying between memory allocations.
Edit: Tegra X1 is cache coherent, so it's safe to assume Switch is too.
McMilhouse
Neo Member
If it was 16nm, instead of 20nm, same horsepower, would it need a fan?
