Daz 3D is part of 
Connect
DAZ Productions, Inc.
7533 S Center View Ct #4664
West Jordan, UT 84084
Licensing Agreement | Terms of Service | Privacy Policy | EULA
© 2024 Daz Productions Inc. All Rights Reserved.
Comments
On the other hand, 3090 users still have 24GB of VRAM, so even if performance is compromised, they'll still be able to use their GPUs on more scenes than owners of smaller cards.
Well I ran the baseline tests wiuth Caustics on and off. Caustics seem to be working because it killed the card performance (doubled the time to render). It was such a big performance hit that I reran the caustics off test again to make sure something else was not causing it. Shutting down now to install my RTX 3060 12GB. More to follow...
System Configuration
System/Motherboard: Gigabyte X570 Aorus PRO WIFI
CPU: AMD Ryzen 7 3700X @ 3.6 (boosts to 4.3)
GPU: EVGA GTX 1070 GAMING @ 1506 MHz
System Memory: G.Skill 32 GB Trident Z Neo DDR-4 3200 memory @ SPEED
OS Drive: Samsung 860 EVO 1TB SSD
Asset Drive: 2 x Western Digital Blue WD20EZRZ 2TB HDD in mirrored array
Operating System: Windows 10 Professional 64 21H1
Nvidia Drivers Version: 456.38 Studio Driver
Daz Studio Version: 4.20.0.17
Optix Prime Acceleration: N/A
Caustic Sampler off
1st run
2022-04-30 13:53:35.177 [INFO] :: Total Rendering Time: 11 minutes 20.70 seconds
2022-04-30 13:59:33.688 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : Device statistics:
2022-04-30 13:59:33.688 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (NVIDIA GeForce GTX 1070): 1800 iterations, 1.138s init, 677.360s render
2nd run
2022-04-30 19:18:28.903 [INFO] :: Total Rendering Time: 11 minutes 21.0 seconds
2022-04-30 19:19:29.399 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (NVIDIA GeForce GTX 1070): 1800 iterations, 1.345s init, 677.414s render
Caustic Sampler On
2022-04-30 18:45:41.243 [INFO] :: Total Rendering Time: 22 minutes 50.76 seconds
2022-04-30 18:47:36.334 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (NVIDIA GeForce GTX 1070): 1800 iterations, 1.146s init, 1367.388s render
So caustics ON still basically doubles the render time. Yet we are seeing some kind of caustic effect with 4.20 even when OFF, which is clearly different from previous versions of Daz.
Perhaps when Richard said that the caustics setting was becoming depreciated, it was because a NEW form of caustics simulation is being done with Iray making the old option pointless. I could be wrong, but regardless of what the deal is, there should be an option for users to disable whatever this new caustics is. It may not be a massive performance hit...but it is a performance hit. I would wager that the vast majority of Daz Iray users never used caustics at all, or cared much about it. Honestly in my opinion caustics can sometimes look more unnatural, like the effect was added in post, even if it was not. With real caustics, you now need to worry about having the correct refraction for every object in your scene, or you may up with something that looks like JJ Abrams Lens Flare.
The same goes for the whole ghost lights disaster. There is a simple fix for how Iray changed ghost lights. In 4.20, the light opacity is linked to the mesh opacity. This does not make sense...is a flame an opaque object? What is the actual physical link between an objects opacity and the light it creates in the real world? It doesn't make any sense. The correct answer is to have TWO opacity values when a mesh surface has light enabled. You have the standard opacity that has always existed, PLUS a new opacity map for the light that is independent from mesh opacity. Bang. Problem solved. You can have a 0 opacity object with 100% visible light. I do not see why this is even a problem.
But I suppose 4.20 does allow you to make volumetric smoke for that 0.5% of the time you want smoke in a scene. That makes all the performance losses and ghost light breaking worth it...hooray. <.<
I would stress to people to keep a backup copy of your old Daz Studio, or try the 4.20 BETA. Do not let your 4.15 or 4.16 go. If you do not keep your old install, you will NOT be able to download it again. It is ENTIRELY on you, the customer, to keep your old installs of Daz Studio as they do not provide downloads to past versions.
All I have to say is WOW. Big difference in render time. Going to go try to render some of my previous scenes that were too slow before.
System Configuration
System/Motherboard: Gigabyte X570 Aorus PRO WIFI
CPU: AMD Ryzen 7 3700X @ 3.6 (boosts to 4.3)
GPU: EVGA RTX 3060 12GB
System Memory: G.Skill 32 GB Trident Z Neo DDR-4 3200 memory @ SPEED
OS Drive: Samsung 860 EVO 1TB SSD
Asset Drive: 2 x Western Digital Blue WD20EZRZ 2TB HDD in mirrored array
Operating System: Windows 10 Professional 64 21H1
Nvidia Drivers Version: 456.38 Studio Driver
Daz Studio Version: 4.20.0.17
Optix Prime Acceleration: N/A
2022-04-30 20:47:30.235 [INFO] :: Total Rendering Time: 4 minutes 31.60 seconds
2022-04-30 20:47:57.894 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (NVIDIA GeForce RTX 3060): 1800 iterations, 9.479s init, 259.889s render
Well that's two of us who upgraded from GTX1070 to RTX3060 this weekend. I also had to upgrade the RAM from 16GB to 32GB. Actual render numbers are almost identical.
System Configuration
System/Motherboard: ASUS H170 Pro Gaming
CPU: Intel(R) Core(TM) i7-6700 CPU @ 3.40GHz
GPU: Asus Dual RTX 3060 OC V2 LHR
System Memory: 32 GB 3200 Mhz DDR4 (MB only supports 2133 Mhz)
OS Drive: ScanDisk SSD Plus 480GB
Asset Drive: Same
Power Supply: FSP Aurum CM AU750M Gold
Operating System: Microsoft Windows 10 Home (x64) Build 19043.1682
Nvidia Drivers Version: Studio 512.59
Daz Studio Version: 4.20.0.17
Optix Prime Acceleration: N/A
Benchmark Results
Total Rendering Time: 4 minutes 22.63 seconds
CUDA device 0 (NVIDIA GeForce RTX 3060): 1800 iterations, 1.653s init, 258.377s render
Iteration Rate: 6.96 iterations per second
Loading Time: 4.253 seconds
Thought I'd add a 2x 3090 benchmark.
System Configuration
System/Motherboard: Gigabyte x299x Aorus Master
CPU: BRAND MODEL @ SPEED/stock Intel i9-10900X @ 3.70GHz
GPU: CUDA device 0 (NVIDIA GeForce RTX 3090) Zotac Trinity v1 stock speed
GPU: CUDA device 1 (NVIDIA GeForce RTX 3090) Gigabyte Aorus Master v2
System Memory: Corsair Dominator 128GB DDR4 @ 2133
OS Drive: Samsung Evo 1 TB SSD
Asset Drive: Corsair Force MP510 series 1920GB NVMe PCIe M.2
Power Supply: Corsair AX1600i
Operating System: Windows 10 Pro Version 21H1 build 19043.1645
Nvidia Drivers Version: 472.47
Daz Studio Version: 4.20.0.3
Benchmark Results
DAZ_STATS Total Rendering Time: 1 minutes 3.84 seconds
IRAY_STATS
CUDA device 0 (NVIDIA GeForce RTX 3090): 830 iterations, 1.161s init, 58.956s render
CUDA device 1 (NVIDIA GeForce RTX 3090): 970 iterations, 1.373s init, 59.173s render
Iteration Rate:
CUDA device 0 14.08 iterations per second
CUDA device 1 16.39 iterations per second
Loading Time: 4.67 seconds
Honestly, I have to say yikes.
Two 3090s took a minute to run this bench in 4.20. To understand just how rough this is, my 3090 all by itself can do this same scene in 92 seconds without any overclock. Two 3090s should be able to hit right about 45-50 seconds in past versions of Daz Studio. But this is clearly not possible in 4.20.
When I run this bench scene with my 3090+3060, I can complete it in 68 seconds in my 4.15. Roughly 8 seconds slower than two 3090s in 4.20. It should not be anywhere near that close, the margin should be double that value.
I sincerely hope you kept your previous install of Daz Studio.
I won't keep blowing up the benchmarking thread after this but one more test to compare 4.15 vs 4.20 on the 3060. Nearly 24 seconds faster on 4.15 and considering this is an animation box, that would mean a quite an improvement in time to do an animated scene. I'll leave 4.15 on this system for now.
I also re-ran the benchmark on my RTX3070 laptop and in 4.16 it does 2 minutes 40.93 seconds and in 4.20 public build it only managed 2 minutes 57.58 seconds.
System Configuration
System/Motherboard: ASUS H170 Pro Gaming
CPU: Intel(R) Core(TM) i7-6700 CPU @ 3.40GHz
GPU: Asus Dual RTX 3060 OC V2 LHR
System Memory: 32 GB 3200 Mhz DDR4 (MB only supports 2133 Mhz)
OS Drive: ScanDisk SSD Plus 480GB
Asset Drive: Same
Power Supply: FSP Aurum CM AU750M Gold
Operating System: Microsoft Windows 10 Home (x64) Build 19043.1682
Nvidia Drivers Version: Studio 512.59
Daz Studio Version: 4.15.0.30
Optix Prime Acceleration: N/A
Benchmark Results
Total Rendering Time: 3 minutes 58.82 seconds
CUDA device 0 (NVIDIA GeForce RTX 3060): 1800 iterations, 3.632s init, 232.572s render
Iteration Rate: 7.74 iterations per second
Loading Time: 6.248 seconds
System Configuration
System/Motherboard: AMD X570
CPU: AMD Ryzen 9 3950X
GPU: 0:RTX 3080 LHR 1740Mhz@0.800V, 1:RTX 3080 non-LHR 1740@0.800V
System Memory: 64GB@3600Mhz (DDR4, 2 Sticks)
OS Drive: 512GB NVME M.2
Asset Drive: 1TB NVME M.2
Power Supply: 1600W Platinum EVGA
Operating System: Windows 10 21H2
Nvidia Drivers Version: 511.79
Daz Studio Version: 4.20.0.17
Peak Power Draw (Total System Max): Idle(pre-render)= 283.500W Peak(post-render): 672.300W (+388.8W)
Benchmark Results
[INFO] :: Total Rendering Time: 1 minutes 8.16 seconds
Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (NVIDIA GeForce RTX 3080): 883 iterations, 1.178s init, 65.326s render
Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 1 (NVIDIA GeForce RTX 3080): 917 iterations, 1.282s init, 64.741s render
Rendering Performance: 1800/65.326 = 27.554 iterations per second
Loading Time: 68.16-65.326= 2.834s
I ran the benchmark in DS 4.20 both with and without the caustic sampler. After running with "OFF", then "ON", I tried "OFF" again in case toggling the switch made a difference, but the render time was the same.
Also ran the same benhcmark with 4.15 with the caustic sampler on - still around 15 seconds faster than 4.20 with caustics on.
System/Motherboard: ASUS X99-S
CPU: Intel i7 5960X @3GHz
System Memory: 32GB KINGSTON HYPER-X PREDATOR QUAD-DDR4
OS Drive: Samsung M.2 SSD 960 EVO 250GB
Asset Drive: 2TB WD CAVIAR BLACK SATA 6 Gb/s, 64MB CACHE (7200rpm)
Operating System: Windows 10 21H2 Build 19044.1645
Nvidia Drivers Version: 511.79
DS 4.20.02
Caustic Sampler set to "OFF":
2022-05-02 19:42:31.987 [INFO] :: Total Rendering Time: 1 minutes 20.65 seconds
2022-05-02 19:42:32.011 [INFO] :: Loaded image: r.png
2022-05-02 19:42:32.056 [INFO] :: Saved image: E:\Users\Joe\AppData\Roaming\DAZ 3D\Studio4 Public Build\temp\RenderAlbumTmp\Render 1.jpg
2022-05-02 19:42:53.054 [INFO] :: Saved image: F:\Windows Folders\Desktop\4_20 caustics off.jpg
2022-05-02 19:42:53.152 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : Device statistics:
2022-05-02 19:42:53.152 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (NVIDIA GeForce RTX 3090): 1221 iterations, 2.416s init, 74.177s render
2022-05-02 19:42:53.152 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 1 (NVIDIA GeForce RTX 2080 Ti): 579 iterations, 2.256s init, 73.879s render
DS 4.20.02
Turned Caustic Sampler "ON":
2022-05-02 19:46:00.841 [INFO] :: Total Rendering Time: 2 minutes 34.37 seconds
2022-05-02 19:46:00.865 [INFO] :: Loaded image: r.png
2022-05-02 19:46:00.907 [INFO] :: Saved image: E:\Users\Joe\AppData\Roaming\DAZ 3D\Studio4 Public Build\temp\RenderAlbumTmp\Render 2.jpg
2022-05-02 19:47:03.875 [INFO] :: Saved image: F:\Windows Folders\Desktop\4_20 caustics on.jpg
2022-05-02 19:47:03.887 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : Device statistics:
2022-05-02 19:47:03.887 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (NVIDIA GeForce RTX 3090): 1260 iterations, 1.470s init, 149.681s render
2022-05-02 19:47:03.887 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 1 (NVIDIA GeForce RTX 2080 Ti): 540 iterations, 1.684s init, 150.455s render
DS 4.15.0.2 Turned Caustic Sampler "ON":
2022-05-02 19:56:27.571 Total Rendering Time: 2 minutes 19.30 seconds
2022-05-02 19:56:27.597 Loaded image r.png
2022-05-02 19:56:27.640 Saved image: E:\Users\Joe\AppData\Roaming\DAZ 3D\Studio4\temp\RenderAlbumTmp\Render 1.jpg
2022-05-02 19:56:40.360 Saved image: F:\Windows Folders\Desktop\4_15.png
2022-05-02 19:56:40.373 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : Device statistics:
2022-05-02 19:56:40.373 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (NVIDIA GeForce RTX 3090): 1281 iterations, 3.096s init, 132.202s render
2022-05-02 19:56:40.373 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 1 (NVIDIA GeForce RTX 2080 Ti): 519 iterations, 3.255s init, 132.681s render
Luckily, I only have 4.20 installed as the beta and 4.15 is still my main rendering version. Might have to stay that way for a while given the speed difference.
Well, that is helpful data. This could be more than just caustics.
Looking at the Release Thread, there is a bullet point that mentions an option for indoor rendering, and they recommend enabling it for indoor scenes.
But I do not see this option. Is there an setting for Guided Sampling somewhere in 4.20? It seems odd they would recommend it but not tell us where it is.
Looking at the Iray Dev Blog, they briefly mention Guided Sampling, and show a comparison shot. The shot with Guided Sampling on has more of a caustic effect, and makes me think of the new caustic effect we are seeing in the bench scene. https://blog.irayrender.com/post/670029739294064640/guided-sampling-iii
Digging deeper I found more info on this Guided Sampler, and it *might* explain what we are seeing here. https://blog.irayrender.com/post/670028436722647040/iray-202110-beta-part-ii
I'll copy/paste the bits here:
That sure does sound like what we might be seeing in our benchmark scene. But instead of having Guided Sampling be an option, it looks like it is just turned on in 4.20. If this is true, then maybe this is not so horrible. But I hope that Daz Studio allows this to be disabled by the user.
If we can disable, then that should return the render speed to normal. And the benchmark save file would need updating to set it correctly. Fingers crossed.
In 4.20.0.17 the 'Guided Sampling' is directly below the 'Caustic Sampler'. Both are off by default.
Well, try turning it on and see if there is any difference, both in render time and in the image. It may be a dummy toggle and be "on" even when disabled in the menu.
This option may have first appeared in one of the last 4.16 versions released. If anyone has that version they could compare them as well.
It is also possible they somehow got the toggle backwards, and off is on, LOL.
Have Beta 4.16.1.43. Still only use that. In last (or almost last) Beta 4.16.*, Guided Sampling is under Caustics. In my tests went with Guided On and Caustics Off. And SSIM available Off and RenderQuality On. Worked best for my 3090 and quick IRAY previews and renders.
Sorry not a scientific Bencher. My interest is more content gen. That said, did those tests quick to optimise my workflow speed.
Do like to read this thread once in a while, especially now with 4.20 being slower for some TBD reason. So thought would post my settings.
Quick footnote, RayDant, your IRAY knowledge has been missed.
With this quick portrait render, guided on took 5 more seconds.
Decided to grab the benchmark one more time, and try it since it has caustics test in it. Had to put different clothes since I deleted the ones the default bench uses. For these two I tried caustics on, once with guided, once without guided. 4 minutes 9.90 seconds with guided on, and 3 minutes 25.70 seconds with guided off.
You can just redownload the benchmark save from the first page and overwrite the one you made, that will restore the scene to where it should be.
It doesn't make sense what is happening here. It seems like they have changed something, since the image has changed a little. The caustic effect at the bottom right of the image is more defined in 4.20, it was much more diffused in every past version of Daz Studio. So obviously something has been changed with how caustics works in 4.20. But it is odd, because nothing in the patch notes mentions this as being default behavior. That is a very important distinction to make.
All of the patch notes mention numerous changes to caustics, but ONLY when caustics or guided sampling is enabled. There is absolutely nothing describing this clear caustic effect when both of those options are off, which is what we are observing.
This has to be where the render times are getting jacked up. A portrait shot does not stress caustics very much, though you may get some caustic action between the eye moisture and tear surfaces. The difference was only 5 seconds. This is pretty close to margin of error, and certainly not like the render times of the bench scene.
BTW, may I ask what skin that is?
Yeah, I just don't have the clothes needed installed in my library anymore lol. Pretty sure the skin is the set from https://www.daz3d.com/mia-hd-for-genesis-8-and-81-female that one, if you mean the closeup one. It's the one I use to render my expression cards with when I need to do that.
I just ran the benchtest scene four times: Default settings, guided sampling on, caustic sampler on and guided sampling on, then back to default.
The guided sampling adds to the render time, but not as significantly as the caustic sampling did. Both together give a significant increase in render time.
There is a difference in the image for each of the first three runs - the bottom of the left hand blue sphere, and under the white sphere at the front right.
There was also something else happening during the guided sampling, as shown in the log file. Not sure how to interpret this though.
2022-05-03 21:33:47.549 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend progr: Received update to 01475 iterations after 79.396s.
2022-05-03 21:33:48.352 Iray [INFO] - IRAY:RENDER :: 1.4 IRAY rend progr: CUDA device 1 (NVIDIA GeForce RTX 2080 Ti): guidance merge/update for: #caches = 65534, #nodes = 84585
2022-05-03 21:33:48.361 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend progr: Received update to 01485 iterations after 80.208s.
2022-05-03 21:33:51.605 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend progr: Received update to 01538 iterations after 83.451s.
2022-05-03 21:33:52.395 Iray [INFO] - IRAY:RENDER :: 1.2 IRAY rend progr: CUDA device 0 (NVIDIA GeForce RTX 3090): guidance merge/update for: #caches = 65534, #nodes = 84592
2022-05-03 21:33:52.397 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend progr: Received update to 01548 iterations after 84.243s.
Test run on DS 4.20.02
System/Motherboard: ASUS X99-S
CPU: Intel i7 5960X @3GHz
System Memory: 32GB KINGSTON HYPER-X PREDATOR QUAD-DDR4
OS Drive: Samsung M.2 SSD 960 EVO 250GB
Asset Drive: 2TB WD CAVIAR BLACK SATA 6 Gb/s, 64MB CACHE (7200rpm)
Operating System: Windows 10 21H2 Build 19044.1645
Nvidia Drivers Version: 511.79
RUN 1 - Default settings
2022-05-03 21:31:34.520 [INFO] :: Total Rendering Time: 1 minutes 20.65 seconds
2022-05-03 21:31:34.544 [INFO] :: Loaded image: r.png
2022-05-03 21:31:34.588 [INFO] :: Saved image: E:\Users\Joe\AppData\Roaming\DAZ 3D\Studio4 Public Build\temp\RenderAlbumTmp\Render 1.jpg
2022-05-03 21:31:45.833 [INFO] :: Saved image: F:\Windows Folders\Desktop\run 1.jpg
2022-05-03 21:31:45.936 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : Device statistics:
2022-05-03 21:31:45.936 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (NVIDIA GeForce RTX 3090): 1220 iterations, 2.359s init, 74.211s render
2022-05-03 21:31:45.937 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 1 (NVIDIA GeForce RTX 2080 Ti): 580 iterations, 2.216s init, 74.206s render
RUN 2 - GS Enabled
2022-05-03 21:34:05.742 [INFO] :: Total Rendering Time: 1 minutes 39.13 seconds
2022-05-03 21:34:05.766 [INFO] :: Loaded image: r.png
2022-05-03 21:34:05.811 [INFO] :: Saved image: E:\Users\Joe\AppData\Roaming\DAZ 3D\Studio4 Public Build\temp\RenderAlbumTmp\Render 2.jpg
2022-05-03 21:34:11.497 [INFO] :: Saved image: F:\Windows Folders\Desktop\run 2 GS.jpg
2022-05-03 21:34:11.512 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : Device statistics:
2022-05-03 21:34:11.512 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (NVIDIA GeForce RTX 3090): 1224 iterations, 1.459s init, 95.018s render
2022-05-03 21:34:11.512 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 1 (NVIDIA GeForce RTX 2080 Ti): 576 iterations, 1.656s init, 95.335s render
RUN 3 - CS and GS
2022-05-03 21:39:01.517 [INFO] :: Total Rendering Time: 3 minutes 6.24 seconds
2022-05-03 21:39:01.541 [INFO] :: Loaded image: r.png
2022-05-03 21:39:01.583 [INFO] :: Saved image: E:\Users\Joe\AppData\Roaming\DAZ 3D\Studio4 Public Build\temp\RenderAlbumTmp\Render 3.jpg
2022-05-03 21:39:24.423 [INFO] :: Saved image: F:\Windows Folders\Desktop\run 3 CS and GS.jpg
2022-05-03 21:39:24.438 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : Device statistics:
2022-05-03 21:39:24.438 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (NVIDIA GeForce RTX 3090): 1221 iterations, 1.477s init, 182.275s render
2022-05-03 21:39:24.438 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 1 (NVIDIA GeForce RTX 2080 Ti): 579 iterations, 1.682s init, 182.362s render
RUN 4 - back to default settings
2022-05-03 21:43:29.998 [INFO] :: Total Rendering Time: 1 minutes 18.38 seconds
2022-05-03 21:43:30.021 [INFO] :: Loaded image: r.png
2022-05-03 21:43:30.066 [INFO] :: Saved image: E:\Users\Joe\AppData\Roaming\DAZ 3D\Studio4 Public Build\temp\RenderAlbumTmp\Render 4.jpg
2022-05-03 21:43:42.570 [INFO] :: Saved image: F:\Windows Folders\Desktop\run 4 back to default.jpg
2022-05-03 21:43:42.580 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : Device statistics:
2022-05-03 21:43:42.580 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (NVIDIA GeForce RTX 3090): 1226 iterations, 1.467s init, 74.697s render
2022-05-03 21:43:42.580 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 1 (NVIDIA GeForce RTX 2080 Ti): 574 iterations, 1.617s init, 74.359s render
@Mart1n71 Thanks for those. I'm thinking about buying a couple of EVGAs now that they are reasonable, but I'm worried about cooling. Did you happend to keep an eye on your temps during the test?
Interesting stuff.
I finally decided to take a shot at the 4.20 beta. I ran a large battery of tests over the past few days. I wont post everything here, I may make another thread. I changed some things around to see more of what the differences are between versions, caustics, and guided sampling.
I found that mesh lights are different between 4.15 and 4.20 even when they both have 100% opacity. The mesh lights are brighter in 4.15, you can see the face of the woman is more blown out by the light. The mesh light is also refracting a little different, and THAT is what we are seeing in the bottom right corner of the image. That white dot is NOT any kind of fake caustic effect at all, it just how the mesh light is passing through the sphere. I placed some diamond OBJs in the scene and found that the light passing through them was also different between the two versions.
So to remove this variable to try and pin down a difference I replaced the glowing ball with a standard pointlight. With this pointlight the renders between the 2 scenes are now identical once again. HOWEVER, there is still a difference in render speed! I had hoped real hard that the render speed was due to the mesh light, so at least some scenes could still render as fast in 4.20. But this is just not the case. 4.20 simply takes longer to render.
I tried to determine if by chance this speed difference was just the short scene getting started, and maybe in a longer scenes this difference be sorted out. So I increased the render size to 1800 by 1800, which is more than twice as many pixels (a lot more). Scaling this up pushed the render times to about 4 minutes even with a 3090+3060. What I found was that YES, the performance difference scales. Not only did it scale, but it was even slightly worse! The performance gap went from 12% at 900p to over 14% at 1800p. While it did not get way worse, it still got worse, and this is the worst possible news. Because the larger and longer your render is, the greater the performance loss might get compared to 4.15. For the people who like making very large scenes...you are getting boned by 4.20. And if you think this is due to temps, my 3090 runs at a flat 63-64 pretty much all the time, and the 3060 also maintains its temp, though slightly higher du to its weaker cooler.
I then tried something totally different. I removed the box room that the scene is inside. I turned on the HDRI, and turned off my new pointlight so the HDRI was the only light. I was hoping that maybe the pointlights and mesh lights were an issue, after all, pointlights still have a mesh. Sadly, the performance with the HDRI was EVEN WORSE. 4.20 rendered a staggering 24% slower than 4.15 with the HDRI lighting the scene with the spheres.
OUCH. That is just terrible. I don't even know what to say. Losing a full quarter of render speed is inexcusable. Just imagine this scaling up...if your render takes 4 hours in 4.15, it might take as much as 5 hours in 4.20, you can seriously lose an ENTIRE HOUR or more because you upgraded software. If you render over night all the time, this may not hurt so much, but if you are rendering and waiting for that to finish to make a new scene or render, you are waiting longer, even though your hardware is the same.
I tried some other scenes. I changed to a i13 classroom that I had, as this scene is sort of like the well known classroom scene benchmark found in Blender. One loads with a camera preset, so I just loaded this up, capped it at 1000 iterations at 1440p and hit render. I rendered with and without caustics. With caustics 4.20 was only 6.5% slower than 4.15, which is much more tolerable. Still, that is a longer time, regardless how small the difference is. It worth pointing out that caustics were not a big part of the scene, it does not have a giant glass ball or mirror surfaces like the benchmark. I can imagine a bar room scene with glass everywhere will be a nightmare.
OK, OK, we have all been testing based on a max iteration count. So I wanted to give 4.20 one more chance. What if 4.20 is doing more work per iteration, and thus the final converged render is just as fast? We have to verify this possibility, too. Maybe all this talk about 4.20 is just a bunch of smoke? <.<
So I did. I went with my pointlight version, with the woman removed, at 1800p. This is your best shot 4.20!
NOPE. 4.20 failed yet again. This time by a full 60 seconds, which was 15% slower than the same exact scene in 4.15.
Alright, what if using multiple GPUs is hurting this? Let's try running that scene with only the 3090.
FAIL. 4.20 managed to mark the beast with a time of 666 seconds, and this was exactly 100 seconds slower than 4.15. That equals 17.5% longer.
I tried all sorts of things, I ran dozens of tests. I tried to find any equal situation between 4.15 and 4.20 where 4.20 could match the speed of 4.15. 4.20 lost EVERY time. The closest it got was with the old classroom scene at 6%. But in these larger tests, you can see the gap increase just about every time. It looks like a 15% difference can be pretty common, and again, it can get to 24%. Who knows, with a truly large scene it might go even higher.
4.20 only offers new volumetrics. That's neat and all, but that just doesn't impress me. 4.20 does do one thing that I have missed since Iray RTX started. It fixes an issue where the textures got duplicated in memory when using the Iray Viewport. If you like using the Iray Viewport, you know what I mean, in recent versions of DS you can quickly run out of memory using the Iray Viewport, and that is because it doesn't share memory with the render. That is finally fixed, so you can use the Iray Viewport in 4.20 in peace.
While Guided Sampling is in 4.20 and helps render caustics faster, this feature is also in 4.16, which renders faster, thus negating this as being any kind of advantage for 4.20.
At any rate, I'll stop ranting, I had to vent a little. I've been at this for days now and I'm ready to move on...with my Daz Studio 4.15. Spread the word. Tell your friends. Shout at anyone who may use Daz, do not upgrade to 4.20, you do so purely at your own risk. Make sure to keep your old versions backed up somewhere. And at every possibility yell at Daz to do something about this mess. They may depend on Nvidia for Iray, but they NEED Iray to be as good as it can be for everybody's benefit, especially Daz. If people use other software to render, they are less inclined to buy Iray specific products from Daz. So while Daz may be more of a content store, that content primarily depends on Iray to sell. It is Daz's best interest to get with Nvidia's Iray dev team and figure out what needs to be done to keep Iray competitive with the fast growing list of GPU based render engines out there. It is not 2015 anymore, GPU rendering is not special.
This is the thing that's so far been pretty much absent to this conversation that I thing really needs emphasizing. Iray is a progressive rendering engine. As opposed to tile based renderers (like the one that drives the commonly used benchmarking tool Cinebench) which rely on highly programmatically taxing algorithms to generate a single comprehensive guess of the values for individual pixels in a given area of a scene before moving permanently on, Iray uses relatively less programmatically taxing algorithms seeded by a psuedo-random number (ignore this detail at your own peril) to generate a much less comprehensive guess of the values for each and every pixel in the scene at once (Compare the image generated by running Cinebench once to a similar looking scene rendered to 1 iteration in Iray - hint: there's no comparison.) It then repeats this process ad nauseum (with a different pseudo-random number value each time to avoid redundancy) averaging the new pixel values with the existing ones to arrive at a (theoretically) constantly increasingly accurate guess at the values for all of those pixels.
Iterations in Iray are not a finite metric of rendering performance in Iray because each and every release of Iray features different versions of the algorithms leading to the generation of individual iterations. Meaning that it is entirely possible for an Iray version X to come about that takes twice as long to reach a given numvber of iterations as an older Iray version Y, despite version X giving a more physically accurate set of pixel values. Remember that Nvidia's goal with Iray is to generat as physically accurate renders as possible - performance is an afterthought. The reason why this thread uses a benchmark scene that is Iteration-limited is because the purpose of this thread is to benchmark rendering hardware against other rendering hardware. NOT benchmarking different versions of Iray against each other (which would require some sort of objective way of judging final render visual accuracy.)
An older version of Iray may reach a certain number of its version of iterations in a shorter amount of time than a more recent one. But barring a comprehensive comparative look into how photorealistic the images rendered are, this time difference is meaningless.
Iterations per second is not a measure of rendering performance. It is a measure of hardware performance while doing rendering. Rendering performance is measured by judging visual quality. Which is honestly beyond the scope of this thread.
Thank you for taking the time to investigate so thoroughly.
For now, I'll use the 4.20 beta if I need the volumetrics, but I'm keeping 4.15 for everything else.
It is absolutely fair to test the software against itself, because this gives us a measure of the software's own efficiency. Efficiency is very important to anyone. Keep in mind that caustics and spectral rendering are OPTIONAL to the user, and they are optional specifically because of the performance hit they cause. The developers have always understood that while accuracy is a goal, you cannot be so accurate that the user needs a super computer to render, that would limit your customer base. The options are there for those who do want them and have either the hardware or time to enable them. But again...they are optional.
At the end of the day Iray is a product, and while accuracy is a nice goal, the true goal is to sell licenses of this product to companies like Daz. It makes no business sense for a company to license a software that their users have difficulty with. Daz Studio was one of the first consumer software to offer a GPU rendering solution back in 2015, and the impact was huge for the company. Nearly every Daz vendor shifted to Iray as fast as they could to take advantage of this new technology which allowed users to render realistic images faster than they could before. Money was to be made.
Nvidia benefits because first they license Iray, but since it is CUDA it only runs on Nvidia. So people who use Daz Studio are far more likely to purchase a Nvidia based GPU, just look at this thread. This certainly effected me, as I owned AMD when I first found Daz back in 2015, and then switched to Nvidia as soon as I could. I bought a 3090 last year, something I absolutely would never have bought if not for Daz Iray. I would have bought a 6900XT otherwise, or maybe nothing new at all. Nvidia sells a good amount of hardware thanks to their CUDA platform. AMD might score some wins with gaming at times, but Nvidia has a huge advantage over AMD in content creation software.
So both Daz Studio an Iray depend on each other. A Daz Studio image is at the top of one of Iray's pages on the Nvidia website, and Daz might be one of the most public faces of Iray around. Whether they want to admit it or not, Iray and Daz Studio are essentially joined at the hip. It makes no business sense to change Iray to be slower because of accuracy, that does not benefit either company. Again, GPU rendering is not unique anymore like it was in 2015.
Additionally, as I mentioned, Iray already has multiple options to make it more accurate, so it would be logical if they had a change to the render calculation, this change would be optional as well. Moreover, the change log makes no mention of any such change, either. It does mention that Guided Sampling will have longer iterations compared to having it off, but this is only brought up with Guided Sampling, and not as default behavior. And considering that they make a point to mention longer iterations with Guided Sampling, I think it would be assumed they would make a point to note any change to the way iterations are done by default.
I rendered a LOT of images in my testing. I looked at the times, and I looked closely at the pictures. I compared both iteration counts and convergence. The lighting from mesh lights is the only difference. Once you remove mesh lights and use standard lights or HDRI, the images are functionally identical. If there is any difference at all, it is so small that no one can perceive it. So if 4.20 takes longer to render and yet has no visible improvement, then there is no argument for accuracy to be made here. The end result is what matters. I encourage anyone to try it themselves and make their own judgements, render your scenes from 4.14-4.16 in 4.20 and compare the images (just not any with mesh or ghost lights).
An iteration in Iray has not really changed since it first started. While new hardware will not operate the old versions, if you have a 900 series card you can pretty much test and verify just how consistant an iteration is across every version of Iray that has released. Even when RTX came along, the work done in an iteration did not really change all that much.
What might change is efficiency. We saw this when 4.14 changed how normal maps were handled. That did not change how iterations worked, it changed the efficiency of normal mapping. This change did not even require a new version of Iray for that matter, the version of Iray was the same! It was Daz themselves that made this change, and it was one of the best things they did with Iray.
So what we are probably seeing in 4.20 is that the introduction of something has harmed the efficiency of the software. IMO, it is not some wonderful push for accuracy, it is simply an optimization issue. I believe one that can be fixed. It might even be something that Daz can help improve, like they did with normal mapping in the past. I can say that normal mapping is not the issue this time, in case anyone might be thinking that normal mapping reverted somehow. I rendered scenes without any normal maps. So sadly it is not that easy this time. But maybe there is something to how Iray is handling a surface or texture that is causing this slowdown.
Doing these tests can matter. The changes to normal mapping came about after I made a thread about how normal maps were so slow. I did a bunch of normal map testing, because I noticed that Daz Original HD models, which do not load with normal maps, were rendering faster than the regular model with normal maps and no HD morph. This made no sense, because normal maps are supposed to be a fast way to fake geometry and save memory. A few months after that, a new version of Daz came along with improvements to normal mapping. Maybe it is coincidence, but if my thread helped bring that issue to Daz's attention, I am happy if it helped. This is why I have been doing this. I am not trying to burn down Daz or Iray, I want to see them improve and be the best they can be. Even when I may sound negative because sometimes these regressions irritate me, it is not because I hate them, it is because I want to see them be better. I want them to succeed. Their success in turn helps all of us. It goes hand in hand.
The nearest thing to a mathematical measure of render quality would be to go back to the convergence criteria. That's assuming they have not changed how they calculate convergence!
Default 95% convergence and Quality of 1:
4.16 reached convergence in 6m11s with 3073 iterations
4.20 reached convergence in 6m50s with 3011 iterations - fewer iterations but took longer. It's doing more work per iteration, but no idea what that extra work is.
Both ways have their pros and cons. The issue with convergence is that Iray only checks for convergence at specific times. As a render gets closer to the target convergence, the checks are more frequent. However what can happen is that if the hardware is very powerful, it is possible to overshoot the convergence target before the check is made. This why we usually see different numbers.
But your numbers are very weird. You got me curious. I did all my convergence tests with pointlights instead of mesh lights, because I was trying to normalize the resulting image. Since mesh lights look different in 4.20, I wanted to make sure that this was not a cause for performance differences.
This is one test with convergence set to 95%. Render size is 1800 by 1800. The ball of light is replaced by a pointlight, and the girl is also removed from the scene. In this test, the difference was only 3 iterations.
4.20 pointlight no girl converged 3090 ONLY 1800p
2022-05-05 16:59:23.218 [INFO] :: Total Rendering Time: 11 minutes 8.85 seconds
2022-05-05 17:46:32.777 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (NVIDIA GeForce RTX 3090): 3312 iterations, 0.290s init, 666.650s render
4.15 pointlight no girl converged 3090 ONLY 1800p
2022-05-05 17:57:12.680 Total Rendering Time: 9 minutes 28.86 seconds
2022-05-05 17:57:24.816 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (NVIDIA GeForce RTX 3090): 3315 iterations, 0.299s init, 566.616s render
So they performed the same iteration count, with 4.20 taking longer. With this test hitting around 10 minutes, only having 3 iterations difference is pretty good.
Your test shows something very different, so I had to see for myself. I went back to the default test and set it to 95% convergence.
4.15 default 95% convergence 3090
2022-05-06 18:33:38.204 Total Rendering Time: 2 minutes 55.24 seconds
2022-05-06 18:33:54.536 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (NVIDIA GeForce RTX 3090): 3296 iterations, 1.711s init, 171.607s render
4.20 default 95% convergence 3090
2022-05-06 18:30:08.453 [INFO] :: Total Rendering Time: 3 minutes 19.36 seconds
2022-05-06 18:30:31.592 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (NVIDIA GeForce RTX 3090): 3248 iterations, 0.864s init, 197.311s render
This time I see the strange iteration difference, yet 4.20 takes longer in spite of doing fewer iterations. This is very odd behavior indeed, and I think this may be unintended. We have people reporting that ghost lights are taking a lot longer to render in 4.20, and these are mesh lights. So this seems to be a trend, mesh lights are doing something weird with convergence. Looking at the two images, I have to say that some spots have more noise in the 4.20 one, not much. It is only 50 iterations out over 3000 iterations, so spotting this difference is not easy. But if you look at the white spot cast in the bottom right corner, the halo around it is very noisy. There is a bit more noise on the ball in front of that white spot, too. What makes this even more strange is that your 3060 saw an even wider gap in iteration count between 4.16 and 4.20. This makes no sense at all. Also, my 3090 ran more than 250 more iterations on its path to 95%. There is so much inconsistancy with this that the data is just usable. But this data shows us that something is up with mesh lights in 4.20.
Regardless of what 4.20 is doing, it is doing it slower than 4.15. If an iteration takes longer, that would be fine if the resulting image at the set convergence still takes the same (or less) amount of time. But that is not the case here. 4.20 is taking longer to hit 95%, and the image at least to my eye is not any better at all. It really seems like mesh lights in general have taken a huge hit in this new Iray. They are pretty much worse in every possible way compared to previous versions. They render slower, present more noise, break ghost lights, and even at 100% opacity they look different. No matter how one feels about ghost lights, the other points cannot be ignored.
So mesh lights are absolutely something that needs a hard look by the Iray Dev Team and Daz, as well as whatever is going on in general. Hopefully we can narrow this down to something so they can fix it.
Incidentally, I found a use for guided sampling! It tries to solve the problem of not yet having refracted rays calculated when pathtracing for reflections.
When shining a beam through a block the reflection in the floor is inline with the spot.
When refracted the spot moves correctly but the reflection remains in the original position.
Switching on Guided Sampling (and leaving it long enough) it starts to find the correct reflection and removes the wrong reflection. Though I took this example from 100 iterations to 5000 and it needed much, much, much more.
Beta - 4.20.1.34
EVGA 3060 RTX default settings, without CPU: 7.05 iterations/s
[INFO] :: Total Rendering Time: 4 minutes 17.42 seconds
[INFO] :: Loaded image: r.png
[INFO] :: Saved image: C:\TEMPDAZ\RenderAlbumTmp\Render 11.jpg
[INFO] - IRAY:RENDER :: 1.0 IRAY rend info : Device statistics:
[INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (NVIDIA GeForce RTX 3060): 1800 iterations, 0.794s init, 255.174s render
EVGA 3060 RTX maximum overclock with CPU: 8.33 iteratons/s
System Configuration
System/Motherboard: ASROCK B450 Steel Legend
CPU: AMD Ryzen 7 3700x @ stock
GPU: PNY RTX A4000 @ stock
System Memory: 48GB RAM DDR4 @ stock
OS Drive: FORCE MP600 512GB
Asset Drive: Innovation IT 2GB
Power Supply: 550 Watt
Operating System: MS Windows 11 (10.0) Professinal 64-bit (Build 22000)
Nvidia Drivers Version: Studio Treiber Version 512.59
Daz Studio Version: 4.20.0.17 64-bit
Benchmark Results
Total Rendering Time: 3 minutes 14.53 seconds
IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (NVIDIA RTX A4000): 1800 iterations, 2.130s init, 190.150s render
Iteration Rate: (1800 / 190.150 ) = 9.466 iterations per second
Loading Time: ((0 + 180 + 14.53) - 190.150 ) = 4.38 seconds
Considering the last generation upgrades, what would be better for iray render... a RTX3080 12GB or a RTX4070 12GB?
RTX3080 12GB specs https://www.techpowerup.com/gpu-specs/geforce-rtx-3080-12-gb.c3834
RTX4070 specs (based on "reliable" rumors) https://www.techpowerup.com/gpu-specs/geforce-rtx-4070.c3890
For comparision, which one is better...
RTX2080S or RTX3070?
RTX2080 or RTX2070S?
GTX1080 or RTX2070?
I'm asking your opinion to decide if I get a 3080 now or wait for 4070.
Thanks