General GPU/testing discussion from benchmark thread

2456718

Comments

  • Robert FreiseRobert Freise Posts: 4,247
    AD said:

    Thank you for this information, but unfortunately the GPU's 1080 Ti and 1070 Ti are no longer available for purchase. Are there any other alternatives that have proven their worth in practical DAZ Studio IRay rendering, but are still available for purchase?
    Thank you for the help.
     

     

    Amazon and Newegg both show to have them 

  • RayDAntRayDAnt Posts: 1,120
    edited April 2019
    AD said:

    Thank you for this information, but unfortunately the GPU's 1080 Ti and 1070 Ti are no longer available for purchase. Are there any other alternatives that have proven their worth in practical DAZ Studio IRay rendering, but are still available for purchase?
    Thank you for the help.
     

     

    Amazon and Newegg both show to have them 

    Just popped over to newegg and wow - the cheapest 2080ti is currently cheaper (at least by US prices) than all but the cheapest 1080ti. That's just bonkers.

     

    Are there any other alternatives that have proven their worth in practical DAZ Studio IRay rendering, but are still available for purchase?

    Pretty much any 10xx or 20xx series Nvidia card will work fine. The key question is how much video memory you'll need for the type of scenes you want to make. A ballpark figure you could use is roughly 2GB per fully outfitted figure plus 2GB for the overall environment. So 2 figures + environment = 6GB mninimum. So anything at or above a GTX 1070 or RTX 2070. As a rule an RTX card will give you singificantly lower/better rendering times than its GTX 10xx counpterpart. But there's nothing a 2070 can do that a 1070 can't do with additional processing time.

    Post edited by RayDAnt on
  • ebergerlyebergerly Posts: 3,255
    edited April 2019

    I'm hoping that, at some point, the 20xx series cards will distance themselves completely from the 10xx series cards in one or more of the following areas:

    1. Make the Iray preview completely realtime, like Blender's Eeevee, even with stuff like "point at" being used (which slows the preview to a crawl). That would be huge to me. Realtime adjustment of depth of field in complex scenes, and so on. I'm hoping they can integrate the game technology of RTX with Iray to get something like that. That would be, for me, a clear reason to spend the big bucks on the 20xx cards.
    2. Realtime cloth simulation. And I mean REAL time, where you can adjust and tweak cloth settings in real time and see immediate draping response. That would be huge. Manipulating the preview and adjusting cloth in realtime would be heaven. I'm not sure if that's the FLEX technology or something else, but one reason I'm kinda biased towards gaming cards like the RTX is that there's so much development towards real time realism in gaming, 

    Other than that, raw render time, for me, is far less of an issue. As is VRAM size. Heck, I have to work very hard to fill up my 1080ti's 11GB of VRAM.

    But being able to get an Eevee-quality preview (apparently also similar to some of the new games) so you know exactly where your final render will be, and being able to manipulate cloth in real time is huge, IMO. Those two things will be my RTX benchmarks going forward.

    But between that and the current prices, I can't even imagine buying a GTX card right now. Although I also won't be surprised if the RTX cards kinda fizzle when all is said and done.   

    Post edited by ebergerly on
  • outrider42outrider42 Posts: 3,679
    edited April 2019

    People all have different priorities, which is why a well rounded discussion is needed. While some people may not care about lots of VRAM, but then you have people like kyotokid who that are extremely focused on VRAM above all else. And then there is the most important factor of all: budget. One person might think $300 is a lot to spend for a GPU, while somebody else is ready to jump at the RTX Titan with no hesitation.

    So it is very difficult to tell any one person that they need to buy this or that. Each of you ultimately need to make your own decision that balances speed, VRAM and budget with the information provided. I often talk about the future because I believe that is a necessary thing to factor. If a new tech is just around the corner, it is only smart to consider it when looking at buying something. 

    I think that one of Daz's "cool new things" that Steve talked about could be some kind of real time mode. The "Iray Viewport" is actually OpenGL, and quite different from the Iray used to render. This is almost a video game like mode as it is. So I can see this happening. It would be even better if users could open this mode in its own window and save the results, like the Spot Render Tool does. This would be quite a game changer indeed. I can see a lot of people using this for final renders if the quality is "good enough". For most people, I would guess raw render time matters quite a bit, that's why the bench thread has been so very active since its creation, and why people talk about the hardware all over the forums. If people can render out in near real time with RTX in Daz Studio, people's brains will explode.

    Dforce is not based on Nvidia. Any change to a real time simulation would be a massive overhaul, and probably not be dforce at all. Personally I'd love to see PhysX used. It is from Nvidia, but it has now gone open source. PhysX can work in real time, and it can work on anything. PhysX can even work for sound. I know, that probably doesn't matter to most, but it would for animators...imagine your animations having physically accurate sound effects. You would not specifically need any RTX card for PhysX, either (just like dforce.)

    Post edited by outrider42 on
  • DAZ_RawbDAZ_Rawb Posts: 817

    So, now that the 4.11.0.335 Daz Studio Pro Beta is out and adds hardware accelleration for the AI denoiser, any good ideas on how to benchmark the denoiser?

  • ebergerlyebergerly Posts: 3,255
    edited April 2019
    DAZ_Rawb said:

    So, now that the 4.11.0.335 Daz Studio Pro Beta is out and adds hardware accelleration for the AI denoiser, any good ideas on how to benchmark the denoiser?

    Wow, that ain't easy...

    I'm guessing you'd have to render an image to perfection, then do the same render with denoising on and somehow figure out how similar the denoised one is to the perfect one as time progresses. Maybe NVIDIA has some cool software that figures out how close an image is to a perfect render or something, but without that it's a bunch of people saying "yeah, that looks real close and I can't tell the difference and it took only 25% of the time of a full render". And getting a bunch of people to agree on what's "real close" is virtually impossible laugh 

    Or, more likely, we'll get some people saying "Dude, that's a game changer !!" and others saying "are you kidding? It looks like junk to me". 

    Post edited by ebergerly on
  • outrider42outrider42 Posts: 3,679
    edited April 2019

    Tensor cores are supported now? Interesting. That's a tough one. I think we would need a scene that is very taxing and long to render. Maybe the Dancer Tutorial scene adapted to Iray, with some lights purposely positioned to cast dark shadows in the corner...the kind of thing Iray can get grainy with.

    Set the render settings to end at either 25% or 50%, and compare this VS a fully converged (or default converged as it is 95%) render.

    What I want to know is if there is a difference in visual quality between using Tensor cores VS pure software denoising. I don't have RTX, so can anybody try testing this out? Doesn't have to be a benchmark, just a visual test.

     

    ebergerly said:
    DAZ_Rawb said:

    So, now that the 4.11.0.335 Daz Studio Pro Beta is out and adds hardware accelleration for the AI denoiser, any good ideas on how to benchmark the denoiser?

    Wow, that ain't easy...

    I'm guessing you'd have to render an image to perfection, then do the same render with denoising on and somehow figure out how similar the denoised one is to the perfect one as time progresses. Maybe NVIDIA has some cool software that figures out how close an image is to a perfect render or something, but without that it's a bunch of people saying "yeah, that looks real close and I can't tell the difference and it took only 25% of the time of a full render". And getting a bunch of people to agree on what's "real close" is virtually impossible laugh 

    Or, more likely, we'll get some people saying "Dude, that's a game changer !!" and others saying "are you kidding? It looks like junk to me". 

    The new denoiser is supposed to be trained by AI, so there is a software element at play. In my own testing, the quality of the image depends on what the image has in it. The denoiser works really well on non organic items. It can be amazing on killing grain in the shadows. Where the denoiser is less exciting is on human skin and some very finely detailed textures. Again, depending on how the scene is set up, the denoiser can wipe out detail in skin making it look flat and smooth. Perhaps good for younger people, less so for older people with weathered skin.

    So that is why you see people saying different things about the denoiser. It all depends on what they do. For some people the denoiser is totally legit and it is genuinely hard to spot the differences between a denoised image VS a fully rendered one. But for others the denoiser is horrible and washes out the details they strive for. Both of these views are correct! Because the denoiser is capable of both good and bad. It is also worth pointing out that letting the render go can help with the clarity when using it.

    So what I want to know is if the new Tensor powered denoiser has better quality than non Tensor. That is more important than the speed. The denoiser is ALREADY pretty fast as it is when it does work. 

    Another thing that many people do not know, hell I didn't know until a week ago, is that denoiser can be turned on and off DURING the render! How many of you knew this???? So you can totally test this without effecting your render. If you don't like it, simply turn it back off before the render finishes.

    If anybody is wondering how to do it, first of all, you need to turn on the denoiser. Seems obvious, but note that there are TWO toggles for turning the denoiser on. One makes the denoiser available, this MUST be on first. The second actually turns the denoiser ON, this can be off. Once you start the render, look on the left side of the render window. There is a very tiny and extremely well hidden notch hidden very well on the left side of the render window. All these freakin years I never noticed this thing. Click on this and a new menu pops out of the side of the render window. Here you can turn the denoiser on and off as you desire. It takes a moment for the new setting to kick in, so be patient.

    Additionally, there are other settings you can tweak while the image renders. You can turn bloom on or off here as well, and even adjust all of the bloom settings. This way you can hone the amount of bloom you desire during the render process, saving a lot of time! Tone mapping is available to alter as well.

    Take note that once the render is finished, you can no longer tweak these settings this way. So you want to get them to where you want before the render is done.

    This is really cool stuff, but holy crap I wish it wasn't so well hidden. These little toggles need to stand out a LOT more so that people can see them. Yes, people can adjust color settings of the GUI, but the vast majority of people choose a theme and stick with it. None of the themes make these toggles stand out much. The toggle looks like it is part of the window design rather than a button you can click! This is the case for a lot of "hidden" buttons and toggles in Daz Studio.

    Post edited by outrider42 on
  • RayDAntRayDAnt Posts: 1,120
    edited April 2019
    DAZ_Rawb said:

    So, now that the 4.11.0.335 Daz Studio Pro Beta is out and adds hardware accelleration for the AI denoiser, any good ideas on how to benchmark the denoiser?

    Off the top of my head:

    1. Take a scene (with a good balance of exposed human skin, textured/untextured inanimate scene objects, and generous lighting), render it (no matter how long it may take) to 95% completion, and save the resulting image/log file data as your ground reference.
    2. Take the same scene and render it to 1, 10, 100, and 1,000 iterations respectively without AI denoising turned on and make sure to save each render/log file data for comparison purposes later.
    3. Repeat step #2 but with AI denoising turned on (and start iteration set to the same number as that run's total iteration count - ie. 1, 10, 100, 1000 consequitively.)
    4. Arrange all the resulting renders in a grid together with the reference image at the top,the non-denoised renders in a column below it on the left, and then the denoised renders below in a column to the right.
    5. Repeat steps 1-4 for a different version of Iray/Daz Studio.
    6. Take the two master graphics that result and subjectively compare them to each other. This is your quality benchmark.
    7. Take the device render time statistic (from "Device statistics" in the log file - NOT "Total Rendering Time") from each run and compare the time differences with vs without. This is your peformance benchmark*.

     

    * You could also do a time for equivalent quality quality with vs without comparison, but doing so would mean needing to subjectively determine (through many test runs and eyeballing) exactly where that quality equivalency happens. Which the above described steps would only begin to determine.

    Post edited by RayDAnt on
  • RayDAntRayDAnt Posts: 1,120

    Another thing that many people do not know, hell I didn't know until a week ago, is that denoiser can be turned on and off DURING the render! How many of you knew this???? So you can totally test this without effecting your render. If you don't like it, simply turn it back off before the render finishes.

    If anybody is wondering how to do it, first of all, you need to turn on the denoiser. Seems obvious, but note that there are TWO toggles for turning the denoiser on. One makes the denoiser available, this MUST be on first. The second actually turns the denoiser ON, this can be off. Once you start the render, look on the left side of the render window. There is a very tiny and extremely well hidden notch hidden very well on the left side of the render window. All these freakin years I never noticed this thing. Click on this and a new menu pops out of the side of the render window. Here you can turn the denoiser on and off as you desire. It takes a moment for the new setting to kick in, so be patient.

    Additionally, there are other settings you can tweak while the image renders. You can turn bloom on or off here as well, and even adjust all of the bloom settings. This way you can hone the amount of bloom you desire during the render process, saving a lot of time! Tone mapping is available to alter as well.

    Take note that once the render is finished, you can no longer tweak these settings this way. So you want to get them to where you want before the render is done.

    This is really cool stuff, but holy crap I wish it wasn't so well hidden. These little toggles need to stand out a LOT more so that people can see them. Yes, people can adjust color settings of the GUI, but the vast majority of people choose a theme and stick with it. None of the themes make these toggles stand out much. The toggle looks like it is part of the window design rather than a button you can click! This is the case for a lot of "hidden" buttons and toggles in Daz Studio.

    Keep in mind - tweaking certain things (such as environment exposure controls) during an active render will result in an iteration 0 reset. You have to play around with things to see which changes have this effect (it's semi-documented in the Iray programmer's manual.)

  • outrider42outrider42 Posts: 3,679
    RayDAnt said:

    Another thing that many people do not know, hell I didn't know until a week ago, is that denoiser can be turned on and off DURING the render! How many of you knew this???? So you can totally test this without effecting your render. If you don't like it, simply turn it back off before the render finishes.

    If anybody is wondering how to do it, first of all, you need to turn on the denoiser. Seems obvious, but note that there are TWO toggles for turning the denoiser on. One makes the denoiser available, this MUST be on first. The second actually turns the denoiser ON, this can be off. Once you start the render, look on the left side of the render window. There is a very tiny and extremely well hidden notch hidden very well on the left side of the render window. All these freakin years I never noticed this thing. Click on this and a new menu pops out of the side of the render window. Here you can turn the denoiser on and off as you desire. It takes a moment for the new setting to kick in, so be patient.

    Additionally, there are other settings you can tweak while the image renders. You can turn bloom on or off here as well, and even adjust all of the bloom settings. This way you can hone the amount of bloom you desire during the render process, saving a lot of time! Tone mapping is available to alter as well.

    Take note that once the render is finished, you can no longer tweak these settings this way. So you want to get them to where you want before the render is done.

    This is really cool stuff, but holy crap I wish it wasn't so well hidden. These little toggles need to stand out a LOT more so that people can see them. Yes, people can adjust color settings of the GUI, but the vast majority of people choose a theme and stick with it. None of the themes make these toggles stand out much. The toggle looks like it is part of the window design rather than a button you can click! This is the case for a lot of "hidden" buttons and toggles in Daz Studio.

    Keep in mind - tweaking certain things (such as environment exposure controls) during an active render will result in an iteration 0 reset. You have to play around with things to see which changes have this effect (it's semi-documented in the Iray programmer's manual.)

    This is just for general knowledge. This would never work as something to benchmark because you cannot go changing values during a bench, that is bad form. But for general use, this is a tip for people who may not be aware this is even possible. While the iteration could resets, the render image itself does not reset, nor does the convergence. The convergence number might behave weird, but it does reset to 0%. The progress tab on the side also reports when these changes are made and they show in the log file.

    I wonder how many users knew this was a feature. Very few people look over the programming manual.

  • IvyIvy Posts: 7,133

    I was given this site by geek squad today ( yea I raised eyebrow in Skepticism )   for some bench testing tools  for real word app bench testing, before i got some new major upgrades on my system Friday ,  I'm having 2 -1080ti's installed in a new external cooling GPU box and maxing out my system memory capabilities to 120 gigs from 64 & having a new Astek CPU cooler put in after what Ramwolf went through with his i thought a good idea change mine out and replace the radiators fans  they were rattling.

    geek squad said I could share the link to the tools its free to the public  .  so if you yall are interested to see how your gpu & cpu measures up with other folks around the world give it a try you can test ever component of your system  https://gpu.userbenchmark.com/Software

    My set up over all score gave me a score of 79% i was a bit disappointed because i thought i was still in good shape. but not surprised with how fast tec is going and everything. surprisingly my 5 year old 6 core  i7 3.8 Gzh CPU got a decent score ot 91%  but this will be the last upgrade for this system . the next move will be a new system.

    Geek guys told me this tool be great for bench testing overclocking gpu's . 

      i can't wait to see my system score after the upgrades are done.

  • outrider42outrider42 Posts: 3,679

    That is a weird test, I am guessing the number is comparing you directly to other benchmarks to get that percentage. 3Dmark offers comparisons like that as well, but it still gives you a numerical score first. I don't think it will compare too well with Iray because the tests are all gaming centered. I don't believe there are any ray tracing elements in them.

    • - CPU tests include: integer, floating and string.
    • - GPU tests include: six 3D game simulations.

    Also, did the test run both GPUs? That would explain why it would be lower. I can't imagine how the 1080ti would be that far down already. There only a scant few cards that can beat the 1080ti in normal gaming situations. This is a list of GPUs that beat the 1080ti in general gaming: 2080ti, Titan RTX, Titan XP, Titan Volta, and 2080 (plus the 2080 only beats it SOMETIMES, BTW.) And that is about it, unless you include some Quadros, which are probably not being used in that benchmark. AMD does not have a single GPU that beats the 1080ti, and all of the GPUs I listed are over $1000 except for the 2080. So there not a huge number of these out there, and it find it hard to believe a 1080ti would only be at 79%.

  • IvyIvy Posts: 7,133
    edited April 2019

    Hi outrider42

    I have not had the upgrade yet.  The Geek squad is installing the new upgrades this Friday 4/19   2 EVGA 1080ti' black super clocks cards  in a external GPU cooling box and 64 more gigs of RAM and astek CPU cooler 1050 watt powers supply etc. .. Geek guys gave me these benchmark tools so I can do a over all system "( PRE-BENCH)" test and then do a test "AFTER" they have done the upgrades,  so i can compare , it should give me some idea how much longer I can go with this system before replacing the whole thing . when i bought this render desk top 5 years ago it cost me $4200 bucks so I rather upgrade this system until it can't be upgraded any longer,  because my next system build will be around $6800.  to me PC aren't cheap.    my benchmark score of 79% is my over all system before test score, the 980ti's i am using now had a score of 84%  So yes I am hoping to hit the scores in the 90% after i get the 2 - 1080TI'S & memory chips upgrades done. and yes there is a Tab  for this test that does focus on gaming  But I guess you didn't notice the tabs for testing  Desk top and workstations.   which give different stress test for bench marking scores. the test also compares your Processor CPU and ram as well as give you a over all PC score of the health f your system which the geek guys said they found helpful to make sure there no issues after major upgrades like i am having done. So if you look closely you'll see its  not just GPU they test. I just posted the GPU tab test.   i thought these bench test tools was pretty unique.  I am sure others people prefer other tools.

    Post edited by Ivy on
  • IvyIvy Posts: 7,133

    I should mention the only reason I am getting these upgrades is I was pretty lucky the other day & I was able to pick  the last 2 new 1080tis off the EVGA store site. for $629 each.  when normally the cheapest i found them were around $1200 each,  that is why I went with the 1080 ti's instead of going with 1  2080 RTX . 

     yea I am lame I have a service contact with geek guys so they fix and install all my PC high end components,persoanlly I rather pay someone that know what they are doing with upgrades liek these  with a service guarantees than do it myself and royally screw up something.laugh

  • KitsumoKitsumo Posts: 1,210

    I don't know how much faith I'd put in that percentage score, it doesn't say what they're comparing it against. I'd recommend Luxmark for testing rendering performance. It uses OpenCL which isn't Iray, but works in a similar way.

    I dont want to bad-mouth the geek squad guys, I'm sure they're fairly competent at what they do, but I doubt they're experts at PBR rendering. 1st rule of tech support: always sound like you know what you're talking about. 

  • IvyIvy Posts: 7,133
    Kitsumo said:

    I don't know how much faith I'd put in that percentage score, it doesn't say what they're comparing it against. I'd recommend Luxmark for testing rendering performance. It uses OpenCL which isn't Iray, but works in a similar way.

    I dont want to bad-mouth the geek squad guys, I'm sure they're fairly competent at what they do, but I doubt they're experts at PBR rendering. 1st rule of tech support: always sound like you know what you're talking about. 

    I am sure you are right but i hire them because i have a high end system that I do not want to screw up the hardware on installing something myself . better to let Best Buy geek squad guys have at it than let me

  • ebergerlyebergerly Posts: 3,255
    RayDAnt said:
    DAZ_Rawb said:

    So, now that the 4.11.0.335 Daz Studio Pro Beta is out and adds hardware accelleration for the AI denoiser, any good ideas on how to benchmark the denoiser?

    Off the top of my head:

    1. Take a scene (with a good balance of exposed human skin, textured/untextured inanimate scene objects, and generous lighting), render it (no matter how long it may take) to 95% completion, and save the resulting image/log file data as your ground reference.
    2. Take the same scene and render it to 1, 10, 100, and 1,000 iterations respectively without AI denoising turned on and make sure to save each render/log file data for comparison purposes later.
    3. Repeat step #2 but with AI denoising turned on (and start iteration set to the same number as that run's total iteration count - ie. 1, 10, 100, 1000 consequitively.)
    4. Arrange all the resulting renders in a grid together with the reference image at the top,the non-denoised renders in a column below it on the left, and then the denoised renders below in a column to the right.
    5. Repeat steps 1-4 for a different version of Iray/Daz Studio.
    6. Take the two master graphics that result and subjectively compare them to each other. This is your quality benchmark.
    7. Take the device render time statistic (from "Device statistics" in the log file - NOT "Total Rendering Time") from each run and compare the time differences with vs without. This is your peformance benchmark*.

     

    * You could also do a time for equivalent quality quality with vs without comparison, but doing so would mean needing to subjectively determine (through many test runs and eyeballing) exactly where that quality equivalency happens. Which the above described steps would only begin to determine.

    Let's see, I count over 20 renders and probably a couple hours of arranging and comparing results in order to get just a single benchmark.

    Yeah, I don't think that will go over too well. laugh

  • RayDAntRayDAnt Posts: 1,120
    ebergerly said:
    RayDAnt said:
    DAZ_Rawb said:

    So, now that the 4.11.0.335 Daz Studio Pro Beta is out and adds hardware accelleration for the AI denoiser, any good ideas on how to benchmark the denoiser?

    Off the top of my head:

    1. Take a scene (with a good balance of exposed human skin, textured/untextured inanimate scene objects, and generous lighting), render it (no matter how long it may take) to 95% completion, and save the resulting image/log file data as your ground reference.
    2. Take the same scene and render it to 1, 10, 100, and 1,000 iterations respectively without AI denoising turned on and make sure to save each render/log file data for comparison purposes later.
    3. Repeat step #2 but with AI denoising turned on (and start iteration set to the same number as that run's total iteration count - ie. 1, 10, 100, 1000 consequitively.)
    4. Arrange all the resulting renders in a grid together with the reference image at the top,the non-denoised renders in a column below it on the left, and then the denoised renders below in a column to the right.
    5. Repeat steps 1-4 for a different version of Iray/Daz Studio.
    6. Take the two master graphics that result and subjectively compare them to each other. This is your quality benchmark.
    7. Take the device render time statistic (from "Device statistics" in the log file - NOT "Total Rendering Time") from each run and compare the time differences with vs without. This is your peformance benchmark*.

     

    * You could also do a time for equivalent quality quality with vs without comparison, but doing so would mean needing to subjectively determine (through many test runs and eyeballing) exactly where that quality equivalency happens. Which the above described steps would only begin to determine.

    Let's see, I count over 20 renders and probably a couple hours of arranging and comparing results in order to get just a single benchmark.

    Yeah, I don't think that will go over too well. laugh

    Useful information comes at the price of patience. And all you need is a single well-performed run of this benchmarking process per capable graphics card to make final judgments (ie. Just 5 people doing this process once to cover the ENTIRE current RTX card lineup.)

  • ebergerlyebergerly Posts: 3,255
    RayDAnt said:

     

    Useful information comes at the price of patience. And all you need is a single well-performed run of this benchmarking process per capable graphics card to make final judgments (ie. Just 5 people doing this process once to cover the ENTIRE current RTX card lineup.)

    Are we sure that AI denoising performance is reasonably consistent across scenes so that a benchmark result is meaningful? I guess I just assumed that the AI algorithm performance might be very variable depending on the contents of the scene/image. I recall from some Blender denoising results the quality varied based on the image contents. My vague recollection was that dark parts of the image where there was little information were very blurry compared to the lighter parts. 

    Has anyone tried it with the latest Iray stuff? 

  • RayDAntRayDAnt Posts: 1,120
    edited April 2019
    ebergerly said:
    RayDAnt said:

     

    Useful information comes at the price of patience. And all you need is a single well-performed run of this benchmarking process per capable graphics card to make final judgments (ie. Just 5 people doing this process once to cover the ENTIRE current RTX card lineup.)

    Are we sure that AI denoising performance is reasonably consistent across scenes so that a benchmark result is meaningful? I guess I just assumed that the AI algorithm performance might be very variable depending on the contents of the scene/image.

    I'd say it's most certainly going to vary widely from scene to scene based on scene content (as well as render resolution.) Which is why the testing methodology I outlined includes steps 4-5. The idea here is to benchmark the behavioral patterns of the ai denoiser across different versions of itself rather than how a single version of it behaves across different scenes, since the former is (a. much more easily carried out due to its limitations on scope and (b. likely MUCH more useful in drawing conclusions about rendering scenes in general rather than rendering just the specific scene(s) tested.

    Has anyone tried it with the latest Iray stuff? 

    I did some preliminary tests with it on one of the benchmarking scenes last night, but stopped after I realized none of the existing scenes has the right sort of composition to really draw meaningful conclusions. There really needs to be a major close-up of non-uniformly textured surfaces like high quality skin/hair detail (the main thing ai denoising tends to butcher) as well as uniformly textured surfaces like walls of buildings (what ai tends to get right) in a scene for any of this to work. And there just don't happen to be any benchmarking scenes of that sort floating around right now (unfortunately the new, ray-tracing oriented benchmarking scene I am currently beta'ing specifically avoids this sort of scene composition because of compatibility concerns with older, lower vram equipped graphics cards.)

    Post edited by RayDAnt on
  • ebergerlyebergerly Posts: 3,255
    edited April 2019
    RayDAnt said:

     

    I'd say it's most certainly going to vary widely from scene to scene based on scene content (as well as render resolution.) Which is why the testing methodology I outlined includes steps 4-5. The idea here is to benchmark the behavioral patterns of the ai denoiser across different versions of itself rather than how a single version of it behaves across different scenes, since the former is (a. much more easily carried out due to its limitations on scope and (b. likely MUCH more useful in drawing conclusions about rendering scenes in general rather than rendering just the specific scenes tested.

    Not sure I follow. If, for example, the denoiser performs much better on brightly lit scenes than on dark scenes, how would testing across different versions give you a usable benchmark? Wouldn't you have to double your 20 renders to make it 40 renders so that you include the light vs dark results? 

    In any case we can speculate all day, but without actually trying it out and getting a feel for it I think we're just doing a bunch of handwaving. 

    Post edited by ebergerly on
  • ebergerlyebergerly Posts: 3,255
    RayDAnt said:

     

    And there just don't happen to be any benchmarking scenes of that sort floating around right now (unfortunately the new, ray-tracing oriented benchmarking scene I am currently beta'ing specifically avoids this sort of scene composition because of compatibility concerns with older, lower vram equipped graphics cards.)

    This is kinda why I've been saying that it's really too early to be thinking about a meaningful benchmark for the RTX stuff. IMO we need to wait for all of this to shake out in coming months. Yeah, we can do an interim thing as long as everyone's aware it's probably not real accurate and will be invalid in the near future. 

    And so far I think we have what we need. It looks like the RTX renders in about 1/2 the time its GTX counterparts do, and much more than that is still not known. 

    Personally, I think it would be nice if someone could give some input on how the RTX handles the realtime Iray preview compared to the GTX's, but that's just me. 

  • bluejauntebluejaunte Posts: 1,859
    ebergerly said:

    Personally, I think it would be nice if someone could give some input on how the RTX handles the realtime Iray preview compared to the GTX's, but that's just me. 

    Nothing different I can see. Iray preview isn't any different to a normal render as far as I know, only that it let's you interact with the scene while the render is running. I certainly didn't notice any more fluid interaction if that's what you were hoping for. It's still molasses as ever. I've gotten into the habit of having the render preview in a separate AUX viewport with the focus being on the normal viewport. That way interaction is quite a bit less laggy. Still not great but better.

  • RayDAntRayDAnt Posts: 1,120
    edited April 2019
    ebergerly said:

    Not sure I follow. If, for example, the denoiser performs much better on brightly lit scenes than on dark scenes, how would testing across different versions give you a usable benchmark?

    Because then you'd have an idea of how effective version A of the denoiser does its thing on a scene in X amount of time vs. how effective version B of the denoiser does its thing on exactly the same scene in Y amount of time. The idea is to benchmark the denoiser itself - not any particular scene because there are an unlimited number of potential scenes compared to a mere handful of denoiser versions. And being generally familair with denoiser behavior is potentially useful in any scene rendering process since everyone has access to each denoiser version. Not so every possible scene.

     

    ebergerly said:

    In any case we can speculate all day, but without actually trying it out and getting a feel for it I think we're just doing a bunch of handwaving. 

    Speculation/brainstorming isn't handwaving. Speculation is the first necessary step in investigating something concretely. And I'm saying that as a professional academic researcher.

    Post edited by RayDAnt on
  • ebergerlyebergerly Posts: 3,255
    edited April 2019
    ebergerly said:

    Personally, I think it would be nice if someone could give some input on how the RTX handles the realtime Iray preview compared to the GTX's, but that's just me. 

    Nothing different I can see. Iray preview isn't any different to a normal render as far as I know, only that it let's you interact with the scene while the render is running. I certainly didn't notice any more fluid interaction if that's what you were hoping for. It's still molasses as ever. I've gotten into the habit of having the render preview in a separate AUX viewport with the focus being on the normal viewport. That way interaction is quite a bit less laggy. Still not great but better.

    Is that with denoising on? I assumed it would greatly help to get a useful Iray preview a lot quicker. Kinda like Blenders new Eevee realtime preview.
    Post edited by ebergerly on
  • TaozTaoz Posts: 9,714
    edited April 2019
    ebergerly said:

    Personally, I think it would be nice if someone could give some input on how the RTX handles the realtime Iray preview compared to the GTX's, but that's just me. 

    Nothing different I can see. Iray preview isn't any different to a normal render as far as I know, only that it let's you interact with the scene while the render is running. I certainly didn't notice any more fluid interaction if that's what you were hoping for. It's still molasses as ever. I've gotten into the habit of having the render preview in a separate AUX viewport with the focus being on the normal viewport. That way interaction is quite a bit less laggy. Still not great but better.

    A little trick that in many cases speeds up rendering, both in preview and separate window: create a render of the scene in a separate window, size doesn't matter, it can be just 50x50 to save memory, and it only needs to render a few seconds when it has started iterating. Keep that window open in the background and things will normally go considerably faster when you make changes, because it skips most of the preprocessing before it starts to iterate.

    Post edited by Taoz on
  • bluejauntebluejaunte Posts: 1,859
    ebergerly said:
    ebergerly said:

    Personally, I think it would be nice if someone could give some input on how the RTX handles the realtime Iray preview compared to the GTX's, but that's just me. 

    Nothing different I can see. Iray preview isn't any different to a normal render as far as I know, only that it let's you interact with the scene while the render is running. I certainly didn't notice any more fluid interaction if that's what you were hoping for. It's still molasses as ever. I've gotten into the habit of having the render preview in a separate AUX viewport with the focus being on the normal viewport. That way interaction is quite a bit less laggy. Still not great but better.

     

    Is that with denoising on? I assumed it would greatly help to get a useful Iray preview a lot quicker. Kinda like Blenders new Eevee realtime preview.

    In my work I render characters 99% of the time, so the denoiser removes all the detail that I test for and is extremely useless to me I'm afraid. 

  • bluejauntebluejaunte Posts: 1,859
    Taoz said:
    ebergerly said:

    Personally, I think it would be nice if someone could give some input on how the RTX handles the realtime Iray preview compared to the GTX's, but that's just me. 

    Nothing different I can see. Iray preview isn't any different to a normal render as far as I know, only that it let's you interact with the scene while the render is running. I certainly didn't notice any more fluid interaction if that's what you were hoping for. It's still molasses as ever. I've gotten into the habit of having the render preview in a separate AUX viewport with the focus being on the normal viewport. That way interaction is quite a bit less laggy. Still not great but better.

    I little trick that in many cases speeds up rendering, both in preview and separate window: create a render of the scene in a separate window, size doesn't matter, it can be just 50x50 to save memory, and it only needs to render a few seconds when it has started iterating. Keep that window open in the background and things will normally go considerably faster when you make changes, because it skips most of the preprocessing before it starts to iterate.

    You mean a separate render window? I'm trying this and can't see a difference.

    With interaction speed I meant for example dragging a morph slider while Iray preview is running. Obviously the denoiser isn't going to help with that either (@previous comment).

  • TaozTaoz Posts: 9,714
    Taoz said:
    ebergerly said:

    Personally, I think it would be nice if someone could give some input on how the RTX handles the realtime Iray preview compared to the GTX's, but that's just me. 

    Nothing different I can see. Iray preview isn't any different to a normal render as far as I know, only that it let's you interact with the scene while the render is running. I certainly didn't notice any more fluid interaction if that's what you were hoping for. It's still molasses as ever. I've gotten into the habit of having the render preview in a separate AUX viewport with the focus being on the normal viewport. That way interaction is quite a bit less laggy. Still not great but better.

    I little trick that in many cases speeds up rendering, both in preview and separate window: create a render of the scene in a separate window, size doesn't matter, it can be just 50x50 to save memory, and it only needs to render a few seconds when it has started iterating. Keep that window open in the background and things will normally go considerably faster when you make changes, because it skips most of the preprocessing before it starts to iterate.

    You mean a separate render window? I'm trying this and can't see a difference.

    With interaction speed I meant for example dragging a morph slider while Iray preview is running. Obviously the denoiser isn't going to help with that either (@previous comment).

    It probably depends on the content and the size of the scene, but it usually works for me if I have a scene that's slow in updating in the preview. Others have confirmed that it works too. Whether it's different with RTX cards I can't tell (I have a GTX 1070).

  • ebergerlyebergerly Posts: 3,255
    edited April 2019
    ebergerly said:
    ebergerly said:

    Personally, I think it would be nice if someone could give some input on how the RTX handles the realtime Iray preview compared to the GTX's, but that's just me. 

    Nothing different I can see. Iray preview isn't any different to a normal render as far as I know, only that it let's you interact with the scene while the render is running. I certainly didn't notice any more fluid interaction if that's what you were hoping for. It's still molasses as ever. I've gotten into the habit of having the render preview in a separate AUX viewport with the focus being on the normal viewport. That way interaction is quite a bit less laggy. Still not great but better.

     

    Is that with denoising on? I assumed it would greatly help to get a useful Iray preview a lot quicker. Kinda like Blenders new Eevee realtime preview.

    In my work I render characters 99% of the time, so the denoiser removes all the detail that I test for and is extremely useless to me I'm afraid. 

    Darn. I was hoping that RTX and denoising could give us Iray previews like this:
    Post edited by ebergerly on
This discussion has been closed.