Nintendo Switch Tech Speculation discussion

[sebbbi]

I think DF are underestimating the massive amount of tile-unfriendly overdraw that the WiiU tuned vegetation - especially grasses - are causing. Drawing grass back to front with no consideration of screen space locality could cause a lot of traffic between rop cache and main memory.

That's a strong possibility. WiiUs 32 MB eDRAM can store the whole 720p back buffer at once (plus all grass textures). EDRAM bandwidth is basically unlimited (enough for maximum fill at fattest format with MSAA) and EDRAM latency is lower than DDR/GDDR. Thus there's minimum penalty of writing randomly around the screen. As you said, commonly used back-to-front sorting causes pretty random screen space locality. No problem for WiiU or Xbox 360 or Xbox One. PS4 has plenty of excess bandwidth for cases like this. But Tegra X1 has only 25.6 GB/s bandwidth and no embedded RAM of any kind.

Maxwell's tiled rasterizer helps a bit in cases like this. Grass patches have low polygon count, so it can buffer many grass patches and then split them to screen space tiles and raster one tile at time (one read & one write per pixel from memory). However the triangle binning buffer size is very limited for this particular scenario, as Zelda has so much grass. The developer should sort the grass in a way that keeps screen local grass blades near each other in the sorted list. This kind of sorting is however much more complex to implement. Macro tiling in software (viewport cull grass/particles for example to 4x2 split smaller frustums = 400x450 each) is easy, and the sorting actually gets cheaper as N is smaller per bin (N log N scaling). But you need to render grass/particles crossing tile edges to multiple frustums, slightly increasing geometry processing cost (shouldn't be any problem on Maxwell).

And, yeah, it's bilinear filtering on Switch even when docked.

I can also see the bilinear filter lines clearly. Digital Foundy incorrectly stated that texture detail is configured differently for 720p and 900p. It's not. This is basic feature of hardware filtering. The gradient calculation is based on UV difference between neighbor pixels. 900p causes gradient to be smaller -> filtering hardware selects a more detailed mip sooner. Filtering hardware always tries to select as close as to 1:1 pixel:texel mapping as possible. Otherwise the textures would look as low resolution at 4K or 1080p as they do at 720p. In modern games bilinear filtering is not common. Trilinear hides the seam. But I still remember this same discussion when Voodoo 2 and Riva TNT were trading blows

 

[Malo]

WTF I didn't make any of those posts you're quoting @sebbbi

 

[Exophase]

Macro tiling in software (viewport cull grass/particles for example to 4x2 split smaller frustums = 400x450 each) is easy, and the sorting actually gets cheaper as N is smaller per bin (N log N scaling). But you need to render grass/particles crossing tile edges to multiple frustums, slightly increasing geometry processing cost (shouldn't be any problem on Maxwell).

Tegra X1's GPU only has 256KB of unified L2 cache, is 400*450 really small enough of a software macro-tile size?

 

[Silent_Buddha]

My comment referred to Nintendo trying to "educate" their customers by telling them dead pixels aren't defects.
You then proceeded to share links with two major companies acknowledging that dead pixels are defects.

Maybe the next time you accuse somebody of getting their "panties in a twist about nothing" you should provide links that counter their argument, not corroborate it.
I mean it's fine by me that you supported my claim.. it's you who won't be taken very seriously.

Maybe you should learn to read as well.

From the Dell Link

The Premium Panel Guarantee is premium within the industry, because it is guaranteed during the term of the limited hardware warranty to have zero pixels stuck in the "on" position — a factor that results in a bright dot on the display.

During LCD manufacturing process, it is not uncommon for one or more sub-pixels to become fixed in an unchanging state. A display with a 1 to 5 fixed sub-pixel is considered normal and within industry standards.

ONLY pixels stuck in the ON position qualify. Pixels stuck in the OFF position or subpixels stick in the ON or OFF position don't qualify.

As well

The Premium Panel Guarantee is available for UltraSharp, Professional and Alienware monitors sold with systems or as stand-alone units purchased direct from Dell, and is included with Dell's standard 3-year Limited Warranty.

Not all of their monitors are covered by the ON pixel warranty.

Apple on the other hand likes to be vague, however.

If you are concerned about pixel anomalies on your display, take your Apple product to an Apple Authorized Service Provider or Apple Retail Store for closer examination. There may be a charge for the evaluation.

They are kind enough to charge you if they so desire. They are so concerned about the consumer's welfare that they don't want to bother letting the customer know what will trigger charges just for inspection of whether there is a defect (that is undefined) that they might take care of.

Sony had the same policy for stuck pixels as Nintendo for their original PSP. I had one with 2 very bright stuck pixels. I sent it in for warranty and it got rejected. The reason? They told me that because the stuck pixels were not in the center of the screen that it was "acceptable" and thus did not qualify for warranty replacement. #$@%@#% Luckily for me, I had got it at a local store, and after getting rejected I took it back to them and got them to replace it.

Regards,
SB

 

[Silent_Buddha]

I haven't heard of dead pixels in absolutely ages. I assumed it wasn't an issue any more.

All TV and display manufacturer's have similar wording to what Nintendo are using.

For example, LG

http://www.lg.com/ca_en/products/documents/LG LCD Monitor Pixel Policy.pdf

Between 4 and 12 Bright (dependant on display resolution) and/or Dark Sub-Pixels are considered acceptable.

Samsung

http://www.samsung.com/dk/support/repairpolicy/Samsung_Monitor_Pixel_Policy.pdf

For monitors, 0-4 Bright stuck pixels are acceptable depending on the size of the monitor. Between 5-8 stuck pixels are acceptable. Between 5-10 stuck sub-pixels are acceptable.

And that's what they consider their "Zero Stuck Pixel Policy" Ummm, only applies to larger monitors (26"+) as well as certain professional models in the 17-19" range. And ONLY to pixels stuck in the ON position. TV's don't qualify for the "Zero Stuck Pixel Policy" only monitors do.

That said. For the most part stuck sub-pixels are extremely rare in A-grade TV panels. Fully stuck pixels are even more rare. These are what generally go into Tier 1 (Sony, Samsung, LG, Panasonic, etc.) displays that are sold to the West.

B-grade panels have more chance for stuck sub-pixels or pixels. These are generally reserved for 2nd or 3rd tier display makers (like my Wasabi Mango monitor which has 2 stuck sub-pixels) and/or used in products destined for 3rd world countries.

Panels are graded by examining X panels from a production batch. Hence why it's possible for a dead pixel or subpixel to slip into an A-grade panel.

For the most part, stuck pixels in A-grade panels are a thing of the past. That said, it still does pop up in very rare circumstances, hence the standard pixel policies of various display makers allowing for X number of stuck pixels or sub-pixels.

I'm assuming that Nintendo are using A-grade panels and are just using the stuck pixel standard policy of whoever they are sourcing their panels from.

Regards,
SB

 

[Shifty Geezer]

Maybe you should learn to read as well.

You've also missed the argument. It's not about policy, but Nintendo claiming the pixels that don't work properly aren't defects. Every panel is supposed to have 100% working pixels. Sometimes they don't. These are defects; unwanted faults of the manufacturing process. For the sake of economics these defective panels are still supplied to buyers as 'good enough', but they are still defective.

Nintendo claiming the defects aren't defects but are part of the intended design of the panels is what Totz is complaining about.

 

[Silent_Buddha]

You've also missed the argument. It's not about policy, but Nintendo claiming the pixels that don't work properly aren't defects. Every panel is supposed to have 100% working pixels. Sometimes they don't. These are defects; unwanted faults of the manufacturing process. For the sake of economics these defective panels are still supplied to buyers as 'good enough', but they are still defective.

Nintendo claiming the defects aren't defects but are part of the intended design of the panels is what Totz is complaining about.

Oops, I hadn't realized that he was just being pedantic.

This then leads me to believe they might be using an LG panel. LG do not consider them defects either. They are just considered either bright or dark pixels/subpixels in LG documentation. Samsung on the other hand uses the defect wording.

Regards,
SB

 

[Arwin]

Yup. This is a console that will be plugged into my TV a lot of the time but when I take it on the go, I'll have my existing library and no compromises on the games I play. That's where the portable PlayStations failed.

I now just use my Vita to remote play on the PS4 wherever I go. We are just now getting unlimited 4G plans for 35 euro a month, so it's stating to become more and more feasible. Biggest missing piece now is being able to do some multiplayer this way.

 

[tongue_of_colicab]

I have no idea if it's an issue with the Switch's panels or not. It's definitely an issue on a number of units, but I have no idea on the proportions of units affected by it.
With so many companies offering "zero dead pixel" policies nowadays, I also find it strange that Nintendo would have such a defensive and rigid position on it. But to be honest it could be a non-issue on the grander scale.

What does bother me is the "this is not a defect" misinformation.




Nintendo doesn't make display panels. The pixel policies they adapt are or should be aligned with the quality control adopted by the display manufacturer for this specific model.
One could think that if they ordered their displays from a manufacturer that guarantees a very high percentage of units without dead pixels, Nintendo shouldn't have to worry about trading in a Switch for every 1000 sales.
Then again, we never know how much of this is related to actual display defects and how much is related to protecting themselves from scams or lawsuits.






From your Dell link:


From your apple link:



My comment referred to Nintendo trying to "educate" their customers by telling them dead pixels aren't defects.
You then proceeded to share links with two major companies acknowledging that dead pixels are defects.

Maybe the next time you accuse somebody of getting their "panties in a twist about nothing" you should provide links that counter their argument, not corroborate it.
I mean it's fine by me that you supported my claim.. it's you who won't be taken very seriously.

Defect or not, all companies have policies allowing for a certain amount of "non perfect" pixels before the whole panel is considered faulty.

So calling it a defect or not, the likes of Dell and Apple are still telling you a defect pixel doesn't mean the whole panel is considered defect.

 

[Deleted member 13524]

Maybe you should learn to read as well.

Oops, I hadn't realized that he was just being pedantic. I'm the one who should learn to read. And while I'm at it I should learn to tone down the flamebait instead of pressing on.

FTFY.

Defect or not, all companies have policies allowing for a certain amount of "non perfect" pixels before the whole panel is considered faulty.

And all companies should determine how many defect pixels or sub-pixels they deem acceptable so the consumers should know whether or not they're eligible for replacements.
Nintendo's current statement on this subject is crap. All they say is "we don't consider this a defect", which also kind of says "if your screen has a number of dead pixels that is below an imaginary number our reps will pull from their asses at the time you complain, you're screwed".

Can we at least agree on this?

 

[Pixel]

In a few days tech Insight (formerly ChipWorks) will publish their teardown of the Switch and is going to explore the issue some people have with the joycon connectivity issues.
Apparently I've heard they will be doing a die shot at some point as well though this isn't mentioned here in the link.

http://techinsights.com/about-techinsights/overview/blog/nintendo-switch-teardown/

 

[Silent_Buddha]

And all companies should determine how many defect pixels or sub-pixels they deem acceptable so the consumers should know whether or not they're eligible for replacements.
Nintendo's current statement on this subject is crap. All they say is "we don't consider this a defect", which also kind of says "if your screen has a number of dead pixels that is below an imaginary number our reps will pull from their asses at the time you complain, you're screwed".

Can we at least agree on this?

Agreed, Apple, Nintendo, and Sony all have rather vague and crap policies with regards to letting the consumer know whether they qualify for a warranty replacement due to defective/stuck pixels.

However, you were being quite pedantic on focusing on the word defect. As all manufacturers have wording similar to Nintendo's. Whether it be called discoloration, defect, or stuck they all refer to the exact same thing. As I pointed out, LG's own documentation refers to defective pixels in a similar way to Nintendo.

[edit] Just ran across this, Tested managed to get a hold of what Apple were using to determine whether a customer could receive service for dead pixels. (http://www.tested.com/tech/1337-we-uncover-the-dead-pixel-policies-for-every-major-lcd-maker/ ) Up to 15 dead pixels was acceptable for larger sized displays. They didn't like the word defect, either. They like calling them anomalies.

Regards,
SB

 

[BRiT]

At least they're not the exploding and catching on fire anomalies.

 

[Exophase]

I haven't been following what people have been saying about this outside this thread, so sorry if this is a moot point now, but is anyone still unsure about the nature of the SoC used in Switch? Because iFixit has a teardown, and unlike the one posted a few pages ago it has very clean shots of the SoC's perimeter showing that it has the same capacitor layout as the leaked prototype shots and Shield TV 2017, and the same part number as the former.

https://www.ifixit.com/Teardown/Nintendo+Switch+Teardown/78263 (see step 11)

So this should fully put to bed any claims that they upgraded the SoC at the last minute to anything more than very slightly different from the Tegra X1 in Shield TV.

 

[Pixel]

The rear of the circuitboard has even more decouplings capacitors and resistors directly underneath the chip. Could you get your boyz to release a better image of the rears of the 2017 Shield TVs underside?

Spoiler

This image is a little to small and the underside of 2017 Shield TV's circuitboard is partially covered by that brace

 

[bunnybug]

I haven't been following what people have been saying about this outside this thread, so sorry if this is a moot point now, but is anyone still unsure about the nature of the SoC used in Switch? Because iFixit has a teardown, and unlike the one posted a few pages ago it has very clean shots of the SoC's perimeter showing that it has the same capacitor layout as the leaked prototype shots and Shield TV 2017, and the same part number as the former.

https://www.ifixit.com/Teardown/Nintendo+Switch+Teardown/78263 (see step 11)

So this should fully put to bed any claims that they upgraded the SoC at the last minute to anything more than very slightly different from the Tegra X1 in Shield TV.

It should have been put to bed when eurogamer gave out the clocks, based on docs from nintendo

 

[Exophase]

The rear of the circuitboard has even more couplings, resistors and capacitors directly underneath the chip. Could you get your boyz to release a better image of the rears of the 2017 Shield TVs underside?

I don't know, but you'd think that could be subject to the board layout and other components, including what external pins are actually connected to anything. This could be partially dictated by functions that Switch is using that Shield TV is not, like the internal display interface.

The decoupling caps on the top are the ones that nVidia included in the TX1's package, presumably because they're needed to be local enough to ensure power/signal integrity for any of the functions they're providing. Maybe nVidia could have made a different package that was better optimized for Switch's specific application but that's not what we're seeing here. If they didn't even change the package what are the odds that they changed the die?

 

[sebbbi]

Tegra X1's GPU only has 256KB of unified L2 cache, is 400*450 really small enough of a software macro-tile size?

I don't know Tegra X1 tile sizes. David Kanter's desktop Maxwell/Pascal tiling experiments (http://www.realworldtech.com/tile-based-rasterization-nvidia-gpus/) is the all info I've got. He doesn't say it explicitly, but his test runs at 1280x720. There's 10 tiles on X direction and a bit over 5 tiles on Y (rgba16). This means that tiles are 128x128 on rgba16. rgba32 results in 64x128 tiles, and rgba8 results in 256x128 tiles. I would guess that Zelda uses 32 bit back buffer format (rgba8 or r10g10b10a2 or r11g11b10f). Could be LDR pipeline, but if it's HDR, then r11g11b10f should be enough with that look.

AMD L2 cache is roughly 2x-3x of all L1 caches combined (16 KB L1 per CU). L2 is inclusive, so a big chunk of it is the same data that is resident in some L1 cache. AMD GPU of this size would likely have only 128 KB of L2 cache. With this in mind, we could estimate that Nvidia could use 128 KB of their L2 cache to tile buffers. That's only a single tile. Of course the binning would still work, but any kind of software macro-tiling wouldn't produce perfect results. Obviously software macro-tiling would still increase locality and reduce the tiling passes (reduced BW cost), as the geometry binning buffers get filled with smaller portion of the screen at a time.

Screen local z-order sorting of overlapping/interleaved quads has worst case global order equal to simple z-order sort. Screen aligned quad depth sorting results in a DAG of dependencies (based on overlap). In the worst case the result is a list (previous always overlaps current and is before it in z-order). But in this case every quad overlaps with previous = pretty good locality. Sorting like this could work. Would be significantly more expensive than standard sorting (as you need to calculate projected BB of each grass patch and lots of BB intersection tests). Of course if you have additional guarantees from the content, you can get better results by doing local z-sort by area (assuming content guarantees no internal overlaps with neighbor areas) and keep these areas adjacent in the global order (as they are guaranteed to have pretty good screen locality). Same technique works well with particle systems (such as explosions).

 

[swaaye]

errr nevermind. Exophase beat me to it.

 

[Goodtwin]

That's a strong possibility. WiiUs 32 MB eDRAM can store the whole 720p back buffer at once (plus all grass textures). EDRAM bandwidth is basically unlimited (enough for maximum fill at fattest format with MSAA) and EDRAM latency is lower than DDR/GDDR. Thus there's minimum penalty of writing randomly around the screen. As you said, commonly used back-to-front sorting causes pretty random screen space locality. No problem for WiiU or Xbox 360 or Xbox One. PS4 has plenty of excess bandwidth for cases like this. But Tegra X1 has only 25.6 GB/s bandwidth and no embedded RAM of any kind.

This actually paints a pretty good picture for Maxwell's bandwidth saving techniques. In portable mode with only 22GB/s of memory bandwidth the Zelda BoTW runs a near locked 30fps. Wii U had a minimum of 35GB/s with the edram, and another 12.8GB/s from the main memory. The fact that the Tegra X1 is able to run Zelda BoTW in 720P with only 22GB/s of bandwidth is pretty impressive. The jump to 900p seems to overwhelm the 25.6GB/s, the double buffer vsync magnifies the issue. I honestly doubt that the frame time is going much over 33ms, seeing as how I have played long sessions where the frame rate stays solid for good chunks of time. The question is would a triple buffered vsync be better or worse? It causes uneven frame pacing that could be perceived as even worse than a drop to 20fps. Its definitely "framey" when it dips, but it is consistent. When these drops occur, they don't seem to last real long. I would say its rare for them to last more than a few seconds, with only a few areas that seemed to have more sustained issues.