I've been looking into why I can tolerate temporal dithering on monitors, but not GPUs. Using samples taken using my capture card I have analysed how many frames it takes a pixel to change. As with all my capture samples, I captured the windows desktop only. To do this, I decode the first frame of my captured footage and turn it into raw RGB values. I then get the next frame, compare it with the original, and the next frame and compare and so on. I record how many frames it takes for each pixel of the first frame to change. Lets call this Time Till Change (TTC). The image below shows a small selection of the TTCs. It does not look very random, particularly that big blob of 1s and 2s on the bottom half. I found that dithering seems to happen in blobs on the screen. Some blobs didn't even dither over 90 frames. A TTC of -1 means no change in colour value. Below is a frequency histogram of the TTCs obtained for the above. If the TTCs were truly random there would be a smooth curve. The peaks and troughs show that enough in aggregate, the dithering method is not great. The temporal dithering produced by GPUs is not particularly random. There are observable patterns. I have no way of testing how the dithering of a monitor works, but for now I presume they do a better job of producing a random dithering field and that is why they affects me less. I repeated this test on both an RX570 and a GTX660 and saw similar patterns. [Edited as a I found a mistake in the image, will probably find a few more if I look long enough] Update 1 When doing the original testing I noticed a lot of the TTCs seemed to correlate with particular colours. To explore if this was the case I captured the grey colour palette below using a GTX660. The image was displayed using google chrome. The resulting TTCs confirmed my suspicion that the temporal dithering frequency does depend on the colours being displayed. The taller the spike, the larger the TTC, ie the more frames it took for the pixel to change colour (dither). The results are fascinating. Some of the colours, including the background colour did not dither at all (tested for 300frames). Other colours had quite long TTCs, some quite short. The Z part of the labels on the graph show the TTC for that pixel. In simple terms: some colours aren't being temporally dithered, some are dithering quickly and consistently, others are dithering over a wide range of frames. Next I looked at spatial dithering - where the colour values do not change with time - but are still dithered. the graph below shows the Red sub-pixel values for the colour palette. Z-labels show the actual value for the pixel selected. It shows the colours which were not temporally dithered, are not spatially dithered either. I use my GTX660 everyday without any discomfort, its the only modern external GPU I can use without issue. I don't think its special, I think I have simply adapted to it over 7 years of use. Some applications do cause me pain when using it though, namely the updated steam library interface and the odd website. I think whats occurring here is the colour palettes in use are producing different dithering frequencies that I'm not used to it. This could also explain why colour settings improve pain for some people. I will explore this next. Update 2 I changed the colour setting for the GTX660 from default to the other options available: RGB Full, RGB Limited, and YCC444. That screen should look familiar to windows Nvidia users. YCC444 and RGB Limited produced no change. Same TTCs, same patterns. RGB Full, was a different matter entirely. All the dark colours disappeared. I viewed it on my monitor to check it wasn't the capture card messing up, still no dark colours. I think this is what's called Black Crush? but I'm not at all familiar with it. Stranger still? No dithering on RGB Full, none at all. Obviously this requires further investigation. I only captured the grey palette with RGB Full, so I'll be trying out more normal images. Update 3 Tried out some other cards. First the GTX550. It used temporal dithering on both full and limited RGB. But neither used a random algorithm. TTCs were all between 1 and 3. Next I tried an RX570. Its dithering was a bit more random than the GTX660, but it was still affected by colour. Below is the TTCs for Limited RGB set to 8 bit, 1080interlaced - I couldn't get the RX570 to send 1080p to the capture card unfortunately. There were some differences when capturing 10 bit, but I think I have posted enough graphs in this thread already. When set to YCC422 there was no dithering at all.

Seagull Good work - this could provide some interesting results if you are able to do side by side comparisons of dithering on a 'good' setup and a 'bad' setup using the same hardware. Since I have a laptop here that I can use for 8 hours a day on Windows 10 1511 with the graphics adapters disabled, but not for more than 30 minutes on build 1909 in any configuration, my capture data could be very helpful to analyse? Did you capture this using an external capture card on a separate machine? I'm guessing screen capture software wouldn't work even if it's 120fps?

si_edgey Its all captured using an a capture card in a second PC, you are right, screen capture software don't work for GPU dithering. Windows is tricky, I've not been able to find any temporal dithering which I can attribute to windows. Whenever I test intel integrated GPUs, I find no temporal dithering. If windows was using temporal dithering, I would always see some temporal dithering, regardless of which GPU I test. I believe people when they say they have a problem with windows, but either its not temporal dithering, or its very hardware dependent. All my testing is done on desktop hardware, so that could be a factor.

Seagull Could I ask what capture card you're using? If it's not ridiculously expensive I could pick up the same card, use the same settings and send you my laptop captures as I have a desktop PC here I can use for recording? That's interesting that you're not finding any temporal dithering in Windows - perhaps it's exclusively an Intel mobile GPU issue?

si_edgey I use BlackMagic Decklink Mini Recorder 4k https://www.blackmagicdesign.com/uk/products/decklink/techspecs/W-DLK-33 I would caution you though, your laptops hardware may not treat its external monitor ports the same as it treats the internal monitor. You might want to make sure you still have a problem when using an external monitor, as that's the only output you'll be able to capture.

Added an update to the original post exploring colour palettes.

Update 2 explores colour settings.

Seagull Confirms the findings of the user "Guzz" in the Nvidia forums who discovered the dithering registry entry. He said full RGB turns off dithering on his PC while limited RGB enables it. https://www.nvidia.com/en-us/geforce/forums/game-ready-drivers/13/288245/is-it-possible-to-port-dithering-from-nvidia-x-ser/2098174/

KM I must have missed that or I'd have tried it sooner. I have an AMD RX570 and a GTX550 here to try next. Unfortunately I have no newer Nvidia cards to hand.

Would be great to know if Chrome or Firefox do something while Full RGB is on. As the current versions give some of us eye strain on known-good cards.

KM On the GTX 660, RGB Full, no dithering with chrome open. Last Update (I think) trying this stuff out with different cards.

GPU Dithering

Seagull

tfouto

I haven't tried it out yet, but I expect it to be fine. If you can, it'll be in the colour settings for your gpu.

tfouto

Seagull I tried on Windows 10 1909 yesterday, and frankly, it seemed no different, but i just used for a couple minutes. My daily pc is other Windows 8.1, which is fine. I think there is subpixel rendering on Windows 10. Because fonts seem fuzzy, and sometimes i find difficult to focus on. I also think there is something different about color temperature or color space that changed with latest updates.

Sunspark

tfouto No subpixel font smoothing in Windows 10. It's greyscale anti-aliasing.

You know, it's funny.. my work managed laptop is a T480 on windows 10 build 1803. Funny thing is, I also have the usb dock it came with, so usb-c to the dock, displayport output from the dock to an old HP displayport to DVI active adapter to my old CCFL monitor, and the backlight looks more stable than on my regular PC which is Win 7, mini-displayport out to active vga adapter to same monitor. The image is fine and I'm confident you all can use it, but there's this subtle background shimmer that isn't there with the T480 setup as-is. One of these days I should look for a mini-displayport to displayport adapter, so I can plug in the same HP to DVI and see how it looks on the Win 7 machine.

I was just now playing with a laptop from 2005, was a good one back then. Amazing how fast it is with a SSD drive in it, even after all these years. Anyway the hardware isn't that healthy anymore unfortunately. The screen isn't that good today. There's this shimmer or fuzz that's hard to quantify, almost a snowy moire effect. I think it might actually be the backlight beginning to fail because in one corner of the screen there's a crack (not on the glass, but behind it) from someone having dropped or crushed it. CCFLs don't last forever unfortunately.

I'm going to slap OpenBSD on the antique laptop tomorrow probably and see how it compares to XP and Linux which I've already viewed on it.

AgentX20

Seagull I changed the colour setting for the GTX660 from default to the other options available: RGB Full, RGB Limited, and YCC444. That screen should look familiar to windows Nvidia users.

Very curious, on my GTX970, I only have Desktop Colour Depth on #3. My connection is DVI so perhaps that has something to do with it.

Seagull

AgentX20

Just tried out a dvi cable (single link) and saw the same thing with my GTX660. With the AMD card I also recall colour setting options changing depending on the connection type.

Lauda89

Seagull Yep i can confirm it. For example, if i use the display port cable i can set 6 bit or 8 bit (there is a BIG difference because with the 8 bit my monitor use the FRC) with HDMI i don't have this option.

AgentX20

So, any conclusions to be drawn here? Is HDMI and a lower/different colour depth worth further trials?

Seagull

AgentX20

All I can really say is that its worth trying different colour formats on the cards you have if you are having a problem. Whilst my gtx 660 does not dither on full RGB, I can't say if it will be the same for other Nvidia cards. But, I personally would be fairly confident that if I bought a new card I'd be able to find a setting that did not dither.

The other, and in my opinion, more important conclusion is that dithering frequency depends upon the colour being displayed. I think this is why people have problems when software updates or changes, steam and chrome for example. The software itself isn't dithering, but if the colours have been changed slightly that will affect how the GPU is dithering and might produce symptoms.

ShivaWind

This is brilliant work!

si_edgey

Fantastic work here @Seagull, thank you for all your experimentations. I've got a GTX770 sitting here that I was going to put in a machine - if I were to send it to you would that be helpful for running some tests? Very happy to do so.

Seagull

si_edgey

You are very welcome. Thanks for the offer of the card, but there isn't anything else I can think of to test for. I've since moved on to biohacking!

si_edgey

Seagull Thanks for the offer of the card, but there isn't anything else I can think of to test for.

I guess I (selfishly?!) would love to know if there's a colour profile for this card that doesn't use dithering, as it appears that every card you have tested has been different. It would be a great to have a database of non-dithering settings for different Nvidia cards. 😀

Seagull I've since moved on to biohacking!

Very curious to hear about this. Keep up the good work and keep us in the loop with your experiments!

Seagull

si_edgey

If you want to buy your own card I can give you my software so you can do these tests yourself.

si_edgey

Funny you should say that - I've just bought an ATEM Mini Pro, and it can capture direct to USB which is quite handy. I'd need to look as to whether it applies any kind of processing but it might even be interesting to test if it can be used to negate the dithering by acting as an active HDMI processor between the computer and the monitor.

If you're able to provide your software I'd really appreciate it and could continue to post dithering tests of different cards as and when I get my hands on them. I'd be tempted even to order cards from Amazon, test them and return them and post the results to try to find a 'safe' list.

JTL

si_edgey If you're able to provide your software I'd really appreciate it and could continue to post dithering tests of different cards as and when I get my hands on them. I'd be tempted even to order cards from Amazon, test them and return them and post the results to try to find a 'safe' list.

@Slacor and I have "our" project too

Setup isn't exactly trivial because it was a) developed on Linux, b) requires a Python runtime, OpenCV and parts of ffmpeg

https://github.com/ledstrain/videodiff

Expect some more features in the future

Seagull

si_edgey

How programming literate are you? My software works, but its written badly, and requires editing the source code to change anything (file name, frame to start and end on etc). I can either give you the code, or the whole MS Visual Studio project file. The later might be easier to work with if you are unsure. The graphs are made in Matlab, which I can give you the scripts for aswell.

Real time removal of dithering is an interesting concept. Thinking about it now, the easiest way to do it will be to drop the colour depth. That way most dithering will be lost as 2/256 blue and 3/256 blue will both become say 0/16 blue. Edge cases like 15/256 and 16/256 that could flicker between 0/16 and 1/16 could be filtered out with a few frames held back to say look for scenarios where: frame 1 = 0/16; frame 2 = 1/16; frame 3 = 0/16, to change frame 2 to 0/16. The latter could be quite computationally demanding though, can't say if you could do that in real time.

si_edgey

Seagull

How programming literate are you? My software works, but its written badly, and requires editing the source code to change anything (file name, frame to start and end on etc). I can either give you the code, or the whole MS Visual Studio project file. The later might be easier to work with if you are unsure. The graphs are made in Matlab, which I can give you the scripts for aswell.

I used to do some programming when I was doing games development so I'm sure I'd be able to pick my way through it, although the simpler the better please - if the Visual Studio file is the simplest for me to get my head around that would be great, along with Matlab scripts - thank you. 😀

JTL @Slacor and I have "our" project too

Any progress on videodiff?

JTL

Seagull How programming literate are you? My software works, but its written badly, and requires editing the source code to change anything (file name, frame to start and end on etc). I can either give you the code, or the whole MS Visual Studio project file. The later might be easier to work with if you are unsure. The graphs are made in Matlab, which I can give you the scripts for aswell.

Why not just gather everything together and post it on GitHub? If there's anything interesting there either myself or Slacor can look into integrating it to "our" project and give you credit.

diop

Seagull Real time removal of dithering is an interesting concept. Thinking about it now, the easiest way to do it will be to drop the colour depth. That way most dithering will be lost as 2/256 blue and 3/256 blue will both become say 0/16 blue. Edge cases like 15/256 and 16/256 that could flicker between 0/16 and 1/16 could be filtered out with a few frames held back to say look for scenarios where: frame 1 = 0/16; frame 2 = 1/16; frame 3 = 0/16, to change frame 2 to 0/16. The latter could be quite computationally demanding though, can't say if you could do that in real time.

I'm on a laptop with the latest ditherig installed. On spatial (default) my desktop background gradients are smooth, immediately when switching to 'Disable all dithering functions' I can see quite visible banding. So ditherig is working in real time. This I imagine is on the driver level but I am not sure if the actual physical output of a GPU adheres to the same rules. It seems as if the video output on modern devices is hard-wired with some dithering algorithm.

Seagull

JTL

That would necessitate learning how to git. There is nothing at all interesting in there, it just decodes with OpenCV, compares the states of each pixel with the prior frame looking for changes, and writes a CSV file for Matlab to graph.

Seagull

diop ditherig is working in real time

This is different. Ditherig is simply telling the GPU to not dither in the first place, rather than removing dithering that's already been applied.

JTL

Seagull

Seagull That would necessitate learning how to git. There is nothing at all interesting in there, it just decodes with OpenCV, compares the states of each pixel with the prior frame looking for changes, and writes a CSV file for Matlab to graph.

Ah okay. I could probably write a custom implementation of that myself.

I was more interested in seeing it to make sure I replicate it precisely, even if it was trivial.

JTL

Seagull That would necessitate learning how to git

I could find you a nice guide or walk you through it if you'd like

Also good to know for Wiki stuff, etc.

« Previous Page