-
Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
A Nintendo Switch has been taken apart
- Thread starter Dystify
- Start date
Pretty sure they did one for Wii U, there is little point in doing a teardown for any other company as they will tell you the exact specs.
Graphics Horse
Member
Aren't the Foxconn clocks for docked? 🤔
I thought the Shield TV could run as fast or faster than those figures.
I may be getting my wires crossed, were there any specific docked and undocked performance claims from the Foxconn guy?
Skittzo0413
Member
MDave did tests on his Shield TV and found out it throttles either the GPU or CPU (or both to a lesser extent) to levels below the Foxconn clocks after a few minutes at max clocks.
The Foxconn leaker made no mention of docked or undocked clocks. In fact, we still have no official word on whether or not clocks change when docked, although Eurogamer's report says they can.
Are you sure? We wouldn't have needed the high res pictures of the chips right here on GAF and the hundreds of posts debating those if DF would have done that.
A teardown does not equal chipworks die shots. And I wasn't sure, I tried to google it and didn't find it so probably not. They made an article about the Iwata Asks that contained pictures of the mcm though.
blu
Wants the largest console games publisher to avoid Nintendo's platforms.
Ok, I didn't pay close attention to that post by z0m3le and missed that generalisation, which was indeed way too broad, but for some categories (read: BRDF model complexity and target dynamic ranges) of pixel shaders the entire shader can get away with fp16. For instance, every single pixel shader in this video, as last-gen as they are, runs at fp16. And don't worry about wasting your points in this thread - while I've seen my share of ISA-level shader optimisations, I'm not currently on the GPU side of the building in the organisation I work at but I still do low-level optimisations on inherently visual stuff where loss of precision is not taken lightly. Suffice to say I wish I could get rid of all the fp64 in our codebase but I can't ; )
True. And yet, that's somewhat orthogonal to the precision topic.
Graphics Horse
Member
Thanks, I thought so. I am curious if the performance/throttling is identical between the 2017 shield TV and the previous one, I think I read that the chip don't look identical?
Damn it, I don't want a broken screen!
The new Shield TV seems to have a new revision of the Tegra X1. I posted the pictures earlier in the thread.
Nothing new, really. Basically it is still a battle between Eurogamer and Foxconn, and we will most likely only know the answer when someone scans the SOC.
BriareosGAF
Member
Are you watching the 8-bit ALU developments in the DNN-on-GPU space? I can't tell if that will thrill or horrify you. I'm just morbidly curious I suppose...
Sure, but the context here was whether half-precision ALU is some sort of fountain of performance equalization. For effective, measurable improvements in throughput, the devil is in the architectural details, and there are no easy answers in real-world (e.g. non-contrived) workloads. If you had a dollar for every time I said "ah, improving this factor of hardware utilization will finally reduce our net wall time", you could buy a nice sandwich and I would still be full of sad/interesting stories. (If you will be at GDC I will buy you said sandwich.)
Ah, true. Thanks.
ShadowFox08
Member
Its like they're rushing everything, which explains the lackluster lineup and info on online content.
Skittzo0413
Member
Based on the pictures in OP we have reasonably identified (like 90%) the RAM as two 2GB modules for a total of 4GB with 25.6GB/s bandwidth.
Beyond that we aren't really any closer to figuring out final specs. I think we determined the SoC in the OP is almost exactly the same size as a Tegra X1, though that doesn't really mean anything beyond any more than 256 CUDA cores is theoretically impossible.
This gives us no indication of 20nm vs 16nm and certainly doesn't tell us anything about CPU cores or clock speeds.
It's also possible that the unit in the OP is not a final retail unit. That's about it I think.
The latest specs are: we have no idea. Could be anywhere from 4x A57s with 3 usable at 1GHz, 256CC at 768MHz docked to 4x A72s at 1.78GHz and 256CC at 921MHz docked. Or anywhere in between. Or the launch clocks are the former but could be raised after launch, potentially up to the latter.
I think the final bandwidth is still up in the air, even with that find, right? They could have added modifications to the board that increase the bandwidth, like the WiiU had.
blu
Wants the largest console games publisher to avoid Nintendo's platforms.
To quote a colleague from the GPU department: 'Guys, we need to figure out something to do in 8-bit, lest we'll soon be using the minority part of the GPU die.'
Oh, definitely. No silver bullets in any of the computational domains I've ever touched upon.
No clue if and who on the team will go to GDC this year, but I'd love to get to SIGGRAPH this year (again, no idea who from the team will attend this year, but chances are not on my side (again), unless I manage to come up with a paper, which is utterly improbable, given my lowly coder status vis-a-vis the math-wiz kids I'm surrounded by).
dampflokfreund
Banned
so tomorrow ends the hardware embargo... can we expect any news on the specs?
Discokuningas
Member
You say like that isn't better graphics. The thing is, games should be made like that. Graphics that work in good resolution and with a good frame rate. To me that is good graphics. Resolution and frame rate shouldn't be sacrificed over "better graphics". At least not frame rate.
Not likely as that's not info that would be in press material. Pretty much have to wait for a full scale teardown where they scan chips etc or a 2017 devkit guide leaks that has all the details
mario_O
Member
It matters because it will look better on a bigger screen, and maybe devs will have room to add a few more things besides resolution. The better question is why do we need that expensive piece of plastic to connect the Switch to a TV.
jts
...hate me...
Because your TV doesn't have a USB-C input and the Switch doesn't have an HDMI port, so you need a dock or a hub to expand the connectivity options of the console and enable that.
As well as enable the usage of common USB-A accessories at home, like ethernet adapters.
I don't mean bumping the resolution doesn't matter, I mean the "real" power of the switch will always lie in its handheld mode as it is the required mode around which everything was built.
The one you get with the switch doesn't cost as much as one bought separately anyway, so expensive is relative. Also it's way more convenient and useful than just a cable or something.
LordOcidax
Member
And the power, usbs?? Also the form factor... Cmon..
mario_O
Member
Nintendo is charging 90 euros for the dock. Not sure how much it adds to the cost of the console.
mario_O
Member
power is also a cable...I get the form factor thing, but for 90 euros??
Far less, packaging a big piece like that and shipping it separately in fewer quantities costs more. Also I imagine retailer cut is likely to be a good part of the price.
Not that Nintendo is above inflating prices, of course.
(Also my personal theory is that they're planning a 4k dock somewhere down the line and want the price difference to reflect positively, but that remains to be seen)
Diablohead
Member
it's 90 euros because they know there is profits to be made on accessories, 3rd parties will come out with one half the price at some point.
Skittzo0413
Member
I think the effective bandwidth is up in the air because we don't know what's on the die, but the maximum bandwidth for the RAM modules we can see is 25.6GB/s as far as I know. I could be wrong there.
The results I got from the Shield TV were that even when locked to a maximum of 1GHz, if the CPU were under load and the GPU was getting pushed hard too, the GPU would fluctuate between 614MHz and 1GHz, averaging about 768MHz most of the time. If the CPU was allowed to go at its maximum, the GPU would get throttled more.
On the topic of 25.6GB/s bandwidth on a 64bit bus;
I thought I would see if my Shield TV could handle a 4K 60fps HDR 10bit video (on my shiny new LG C6 TV), and it handles it perfectly. Would that mean anything useful in how the X1 is optimised well for high bandwidth usage?
LordOfChaos
Member
Not really though I appreciate the effort, if you think about a 4K blu ray specification being 50GB single layer, 66GB dual-layer and 100GB triple-layer each with 82Mbit/s, 108Mbit/s and 128Mbit/s data read speeds, then even the modest 25GB/s memory bandwidth is magnitudes away from that scale. Playing it is more hardware decoder and/or CPU bound. Doesn't tell us much about gaming or TX1s texture compression or anything of that nature, as you could transfer a whole blu ray in two seconds on internal memory bandwidth
25GB/s eh, so we're about standard for mobile, I was rooting for a 128 bit bus to alleviate that long standing mobile bottleneck (iPad Pro 12" does this). Well, now to see if there's any fancy on-die SRAM blocks by the GPU.
Discokuningas
Member
I don't think games will be made for handheld mode paticularly. Handheld and tv goes hand in hand because of the power difference. So actually developers don't have to target some specs and then port to other. It's just resolution difference. Not saying it's that easy but you get what I mean. And I still don't really get what you are saying. Because still to me resolution is as big part of graphics as anything else. It's improvement as much as better lightning or more polygons would be. And developers actually have that choice if they want to. So tv mode isn't gimped by handheld mode. You can see it both ways. Handheld mode that has been boosted for TV or a TV mode that has been downgraded for handheld. But in the end the automatic power difference is just made for the both modes. So you don't have to sacrifice one or another. What I'm trying to say is that you would probably get pretty much the same results without the handheld mode being an option.
Clock speed doesn't only just affect pixel-related throughput. There's still the front end of the GPU to consider i.e. geometry/triangle setup.
Graphics Horse
Member
Thanks. Presumably this is 4 cores active too? I'm wondering if foxconn was doing some stress test where possibility of throttling was disabled, I guess they give retail chips at least a bit of leeway to avoid melting them too often in the real world.
Curious if these could have been made to hit a higher performance bar (Nvidia making shield TV 2017 from binned switch chips?). But I don't really know what I'm talking about 🐵
lwilliams3
Member
When it comes to Foxconn vs Eurogamer's number, the GPU clockspeeds difference is only 20%. It will not be that notable to the end user. The difference in the CPU clockspeed and chipset, however, is a 75% difference or higher if we are discussing A57s vs A72s.
The CPU speed doesn't change between docked and undock mode, so that is one reason why this remains as a hot topic.
I believe USB-C docking stations are relatively new technology, so they are generally not that cheap to get at this time.
Right. So, it is possible that there will be upgraded/downgrades polygon models and/or viewable geometry between modes, correct? It is possible that we may be seeing something like that with Zelda, but we need to look at the difference between the final versions.
Graphics Horse
Member
This is one of the things I'm wondering is still true, was this only confirmed via Eurogamer?
Skittzo0413
Member
Eurogamer is actually still the only source who has said there are different clock speeds between docked and undocked to begin with. I still think there's a (admittedly small) chance that there is no clock speed change at all, which would explain why titles like Zelda and Splatoon 2 are not 1080p when docked.
But yeah EG is the only source saying the CPU clock will not change when docked/undocked.
forknknife
Banned
I wish I would understand what you guys discuss in this thread 😣
I keep reading it though lol
I keep reading it though lol
lwilliams3
Member
Actually, that is true. We don't have official confirmation of that. Changing the clockspeed between modes could lead to issue with game logic, physics, etc, so it would make sense for it not to change.
mario_O
Member
wut? you're saying both things at the same time.
Eurogamer said CPU clocks won't change, always at 1 ghz. And GPU clocks will change from 307.2 to 768MHz docked.
Skittzo0413
Member
Sorry, I didn't specify: Eurogamer is still the only source we have that actually says the GPU clocks will change between docked mode and undocked mode, meaning they are the only source which says anything at all changes when docking the Switch.
Which makes me still a bit wary of the possibility that this isn't going to be the case for the final hardware, even though we all treat it as though it is.
lwilliams3
Member
Yeah, I seriously doubt that. The patent for the Switch mentions that the system fan is faster in docks mode, and that wouldn't seem necessary if the specs are the same between modes. One of the source from Eurogamer also stated that the difference between docked and Undocked was like making a game with two difference specs. If this isn't true, Eurogamer's info is very wrong, and I don't think anything we know about the system so far discredits their info yet.
Graphics Horse
Member
Yes but could at least be an optional mode for more experienced teams who might need it, boost mode on PSpro (which iirc digital foundry says boosts CPU clock too) sounds pretty successful without devs having had foreknowledge of it.