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
@pandapenguingames that's awesome! Any chance you could post your results for the benchmarks two more times (2. with just a single 2080Ti active for rendering, and then (3. with both cards active but no NVLink enabled. That way we can see how all the different scenarios measure up in terms of performance.
Going by previous dual 2080ti posts, it doesn't look like much of a penalty at all for Nvlink. That's quite interesting indeed.
I would love to know how much VRAM is available. Does it report this in the help log? Have you been able to render a scene normally too large for a single 11GB card?
@outrider42 see this post. The answer to both your questions is already there.
Maybe I am overlooking it but I don't see where it says how much VRAM is available between the two cards.
Oops - my mistake. Yeah, don't see any mention of amount free (total - used from GPU-Z?) reported there either.
NVLink enabled!
System Configuration
System/Motherboard: ASRock X299 Extreme4
CPU: Intel i7-7800X @ Stock
GPU: GPU1 INNO3D RTX2080TI iCHILL Frostbite @ stock
System Memory: 64GB Crucial Ballistix Elite DDR4 @ 2400
OS Drive: Samsung 960 EVO 250GB NVMe SSD
Asset Drive: 2x Crucial MX500 480GB (JBOD)
Operating System: Windows 10 Pro x64 Build 1909
Nvidia Drivers Version: SRD 442.92 WDDM
Daz Studio Version: 4.12.1.117
Optix Prime Acceleration: N/A
Benchmark Results
DAZ_STATS
2020-06-06 10:49:48.734 Finished Rendering
2020-06-06 10:49:48.786 Total Rendering Time: 4 minutes 24.91 seconds
IRAY_STATS
2020-06-06 10:50:09.158 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : Device statistics:
2020-06-06 10:50:09.158 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (GeForce RTX 2080 Ti): 1800 iterations, 2.763s init, 255.506s render
Rendering Performance: (1800 / 255.506) = 7.045 iterations per second
Loading Time: ((0 * 3600 + 4 * 60 + 24.91) - 255.503) = 9.407 seconds
NVLink disabled!
System Configuration
System/Motherboard: ASRock X299 Extreme4
CPU: Intel i7-7800X @ Stock
GPU: GPU1 INNO3D RTX2080TI iCHILL Frostbite @ stock
System Memory: 64GB Crucial Ballistix Elite DDR4 @ 2400
OS Drive: Samsung 960 EVO 250GB NVMe SSD
Asset Drive: 2x Crucial MX500 480GB (JBOD)
Operating System: Windows 10 Pro x64 Build 1909
Nvidia Drivers Version: SRD 442.92 WDDM
Daz Studio Version: 4.12.1.117
Optix Prime Acceleration: N/A
Benchmark Results
DAZ_STATS
2020-06-06 11:08:35.866 Finished Rendering
2020-06-06 11:08:35.905 Total Rendering Time: 4 minutes 21.19 seconds
IRAY_STATS
2020-06-06 11:09:02.918 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : Device statistics:
2020-06-06 11:09:02.918 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 1 (GeForce RTX 2080 Ti): 1800 iterations, 2.727s init, 252.256s render
Rendering Performance: (1800 / 252.256) = 7.136 iterations per second
Loading Time: ((0 * 3600 + 4 * 60 + 21.19) - 252.256) = 8.934 seconds
NVLink disabled!
System Configuration
System/Motherboard: ASRock X299 Extreme4
CPU: Intel i7-7800X @ Stock
GPU: GPU1 INNO3D RTX2080TI iCHILL Frostbite @ stock, GPU2 INNO3D RTX2080TI iCHILL Frostbite @ stock
System Memory: 64GB Crucial Ballistix Elite DDR4 @ 2400
OS Drive: Samsung 960 EVO 250GB NVMe SSD
Asset Drive: 2x Crucial MX500 480GB (JBOD)
Operating System: Windows 10 Pro x64 Build 1909
Nvidia Drivers Version: SRD 442.92 WDDM
Daz Studio Version: 4.12.1.117
Optix Prime Acceleration: N/A
Benchmark Results
DAZ_STATS
2020-06-06 11:16:25.821 Finished Rendering
2020-06-06 11:16:25.857 Total Rendering Time: 2 minutes 19.65 seconds
IRAY_STATS
2020-06-06 11:16:40.572 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : Device statistics:
2020-06-06 11:16:40.572 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 1 (GeForce RTX 2080 Ti): 894 iterations, 2.974s init, 130.255s render
2020-06-06 11:16:40.572 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (GeForce RTX 2080 Ti): 906 iterations, 2.840s init, 130.918s render
Rendering Performance: [(894 + 906) / 130.918] = 13.749 iterations per second
Loading Time: [(0 * 3600 + 2 * 60 + 19.65)] - 130.918 = 8.732 seconds
so regarding the questions about the available VRAM, the only lines I could find in the DAZ log regarding this:
NVLink disabled:
NVLink enabled:
so it's basically reporting 9GB ready for use in both cases, just that with NVLink you can evenly spread out the texture memory consumption across both GPUs
and to get the "just switch to TCC then" argument off the table right away: so far, NVLink only establishes in WDDM and on top of that, TCC can't be enabled on GeForce cards in Windows anyway
Thank you kindly for taking the time to post. We finally have answers to some lomg standing questions about Nvlink. This is great information.
That's impressive, the difference in performance between having Nvlink on vs off is so small there is little reason to disable it for small scenes. That's something I've been wondering.
I'm guessing the theoretical VRAM limit must be 19 GB, minus the data that has to be duplicated. For most users, texture data is the bulk of their scene, so this still a big win all around!
Now I wonder how the 2080 and 2070 Super hold up. These two support Nvlink as well, but their connection is a little slower. It would be interesting to know if that effects the performance. The price of the 2070 Super makes it quite appealing in this regard. Two 2080 Supers should beat a single 2080ti AND offer more VRAM as well while costing roughly the same. The chart has two 2060's on it, which match up with a 2080ti, so it stands to reason that two 2070 Supers should easily beat a 2080ti.
This also changes the conversation around upgrading from a 2080ti. Conventional thought would be to buy the future 3080ti when it releases. But Nvlink could change that since you can effectively "upgrade" your VRAM capacity with a 2nd 2080ti in Nvlink mode. You can't do that if you buy a 3080ti, though a 3080ti+2080ti should be faster. It now becomes a decision as to whether you want the extra speed or the extra VRAM.
@pandapenguingames Just to clarify, as I think the red hightlighted NVLink tags may be mis-labeled... the first 2 configurations shown use a single CUDA device (i.e. 1 x RTX 2080 Ti) and therefore should be labeled "NVLink disabled" (note: the first example you give with only CUDA device 0 has the red highlight saying "NVLink enabled"). These single cards take about 9-seconds to load the scene and then about 4.5-minutes to render the scene.
The final configuration has both CUDA device 1 and CUDA device 0 (i.e. 2 x RTX 2080 Ti), but appears to be mis-labeled as "NVLink disabled". These dual cards (i.e. NVLink enabled) cards take close to 9-seconds to load, but take about 2-minutes less time to render the same scene, that would be nearly twice as fast (about 1.9 x) as a single card... correct?
You can compare PPG's marks to the other dual 2080ti configurations on the first page and see it gets nearly the same results as those non Nvlink systems. So if anybody happens to have two 2080tis, it looks like a no brainer to buy that Nvlink connector and get some more VRAM capacity. Iray has always scaled pretty well, so you've always been able to get that kind of performance when doubling up GPUs before Nvlink came along.
For anyone with a single 2080ti, your options have grown. The 3080 will be releasing this year, without any doubt. You can wait until then to buy a new card or look at buying a 2080ti at a reduced price. I am quite confident in saying that the 2080ti will drop in price pretty rapidly once the 3080 and 3080ti arrive. Also, AMD has confirmed their new lineup will launch before the new consoles do. That means we could be looking at September or October, which has long been rumored a target date. It could even be August, but that may be pushing it. Some people believe the top AMD card might actually dethrone the 2080ti, which could cause a major price shift regardless of when the 3080ti arrives. So much hardware is coming this year its crazy.
It's all labeled properly. @pandapenguingames has so far posted numbers for 4 different NVLink related benchmark scenarios. Consisting of:
The middle two use cases may not be very practical, but are perfectly valid. And useful to know about at this early stage of performance discovery.
@RayDAnt, thanks for clarifying. Those middle two benchmark tests with only a single card used threw me off. I found the most interesting part to be the first test with both cards and NVLink enabled was a little slower in render time, slower in loading time, and had slightly less iterations per second than the fourth test with both cards and NVLink disabled. From those benchmarks, it appears that having 2 cards with NVLink disabled is slighty faster.
this test was less about speed and more about memory pooling for me, I only did the benchmark on request of others
I don't care about render speed if I can't fit the scene in my VRAM and since the NVLink is not adding significantly much to the render times, I will run on NVLinked cards from now on
@pandapenguingames Ok, I see the benefits now… the main benefit of the 2 cards with NVLink Enabled is to be able to load more texture data into the GPU memory (i.e a larger scene), while only sacrificing very little in regards to speed.
I think the owners of any NVLink capable card has to give serious thought to adding a matching card and an NVLink bridge rather than upgrqading, unless the Ampere cards carry substantially more VRAM at the same price points.
The 2070 Super might just be the sweet spot for this. 2 2070 S and the bridge is $1100 which is right at the same price point as a 2080 ti. But the pair has 5120 CUDA and 14+ gb of VRAM compared to the 2080ti's 4352 CUDA and 11Gb of VRAM.
Yes, that was expected. The actual numbers were of most interest here. Nvidia explained at the very start during the launch of Turing that Nvlink would incur a performance hit compared to that of a single card, or for Iray, running multiple cards without the Nvlink. You have to consider what is happening here. The CPU needs to feed data two separate cards, and it has to decide which data is duplicated or not and get everything straight before a render. That's going to take a little longer. The actual render the cards have to talk to each over over the Nvlink because they share that data. So as each card works, it needs to be able to address data over the Nvlink bridge, and while Nvlink is fast, it is still not as fast as the GDDR6 that is located on the card right beside the GPU chip. So it takes a fraction of a second longer to retrieve this data, and that adds up over the course of render.
Other render engines that support Nvlink have all shown that performance drops a little compared to running without Nvlink. The big advantage to Nvlink that we have all been looking forward to is pooling VRAM to be able to render larger scenes. I posted Vray Nvlink tests some time ago, which showed I believe a 5% or so performance hit with Nvlink. Even 5% is pretty small.
However, just going by PPG's numbers, it looks like the performance hit for Iray using Nvlink is quite small. Maybe its because the bench scene renders so fast, but its only 3 seconds longer for the Nvlink devices to render.
I have been quite hard on Iray, but it looks like they have done a solid job of getting Nvlink. It took 2 freaking years, but they did it.
That depends. If the 3080 or 3080ti are significantly faster than the 2080ti, which I expect them to be, then people who want pure speed may look at adding the 3080ti to the 2080ti. There is little doubt that this combo will be a lot faster than two 2080tis. And if VRAM is not a huge concern, I can see people doing that instead.
But if VRAM is a big concern, then the choice is pretty easy to buy a second 2080ti. Unless of course they have the cash to buy two 3080tis or 3080s and use Nvlink on them.
Certainly the 2070 Supers are very interesting. For anyone kind of on a budget, that's a great deal. But nobody has actually tested it yet, so I cannot advise people to jump on it. We do know that Nvlink works on the 2070 Super, games have been tested with it, and at least in games, the combo beats a single 2080ti pretty easily, but not by that huge a margin. Iray would probably scale better, but we don't know for certain.
The trouble is that this all comes out right at the end of the Turing generation. Most people have probably forgot about Nvlink at this point, or given up on it. A lot of people here didn't think VRAM pooling would happen, and if they don't check the forums a lot, they might never find out.
AHh glad its not just me or something wrong with my new GPU then. I've asked Daz for version 4.12.0.86 as that is what's currently on my old PC and doesn't have the newly introduced animation bugs but I can't find a way to transfer it to my new PC, I unfortunately had "delete package" ticked on DIM when I initially installed it. Its been nearly a week and no response from them yet though.
Go to "C:\Program Files\DAZ 3D" on the machine with the older relase and copy the folder "DAZStudio4" onto a flash drive (I suggest also compressing/backing it up to a more permanent location with a name like "DAZ Studio Pro 4.12.0.086.zip" or something for future use.) Plug the flash drive into the machine with the newer relase and overwrite the "DAZStudio4" folder found in the same place with the folder copy you just made.
Alternatively, you can also just copy that folder to any location on the other machine and run the previous DS version simply by double-clicking "DAZStudio.exe" found inside the folder. That way you don't have to worry about possibly breaking the newer installation or DIM overwriting it without your knowledge.
Where can I get the B25STTop shirt and textures used in the benchmark scene?
Everything used in the benchmarking scene is included with the DS base files.
@RayDAnt
Would you mind if I share a conversion of your scene I did for Blender? Someone was asking. It isn't possible to get a good comparrison of times; I merely wanted an idea to judge if it was worth it to me; besides the character was transferred using Diffeimorphic, and that has very custom shaders, designed to replicate the look of Iray in Studio.
two wildly different 'itteration' renders and one ran through the denoiser (Intel's).
Sure - share away!
System Configuration
System/Motherboard: MacPro5,1 (Mid 2010)
CPU: 2x 6-Core Intel Xeon @ 2.66 GHz (Westmere)
GPU: AMD Radeon HD 7950 (3072 MB)
System Memory: 32 GB 1333 MHz DDR3 ECC
OS Drive: Hitachi HDS722020ALA330 2 TB
Asset Drive: Hitachi HDS722020ALA330 2 TB (not system drive)
Operating System: Mac OS X 10.11.6
Nvidia Drivers Version: N/A
Daz Studio Version: 4.12.1.086 64-bit
Optix Prime Acceleration: N/A
Benchmark Results
2020-06-09 10:09:30.721 Finished Rendering
2020-06-09 10:09:31.129 Total Rendering Time: 1 hours 6 minutes 35.33 seconds
2020-06-09 16:01:23.633 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : Device statistics:
2020-06-09 16:01:23.633 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CPU: 1800 iterations, 6.004s init, 3982.505s render
Iteration rate: 0.45197683367629 iteration per second
Loading Time: 3995.66 seconds
Note: Did several tests with and without running other programs and multitasking. There was little to no difference in benchmarks with multiple apps running (3 web browsers, Mail, Text editor, utility apps)
System/Motherboard: GIGABYTE B450 AORUS PRO WIFI-CF
CPU: AMD Ryzen 5 2600X @ 4.0 ghz
GPU: MSI GTX 1070 Armor OC/Aero OC @ 1648mhz clock 2230mhz memory, GIGABYTE GeForce GTX 960 OC 4GB @ 1276mhz clock 1810mhz memory MSI RTX 2080 Super Gaming X Trio
System Memory: Corsairw DDR4-3200 8gb x2 gskill DDR4-3200 8gb x2
OS Drive: Samsung 960 EVO 250gb
Asset Drive: Seagate 2TB 7200 rpm
Operating System: Win10 Enterprise 18363
Nvidia Drivers Version: 442.19
Daz Studio Version: 4.12.2.6
So TheKD also lost time, with the device render time going from 283 seconds with driver 436 and 4.12.0.067 to 313 seconds with driver 442 and 4.12.2.6. Almost exactly 30 seconds slower. The loading time is actually faster, but the overall time is still much slower.
This trend is disturbing.
Hi, here my stats
System Configuration
System/Motherboard: ASUSTek PRIME X370-PRO
CPU: AMD Ryzen 7 1700 @ 3700 MHz
GPU: GPU1 Manli GeForce RTX 2060 Super 8GB Gallardo @ stock, GPU2 Manli GeForce RTX 2060 Super 8GB Gallardo @ stock
System Memory: G.Skill DDR4 8GB x4 (32GB total) @ 1600 MHz
OS Drive: MyDigitalSSD bpx NVMe 118GB
Asset Drive: SanDisk SDSSDHII480G 480GB
Operating System: Windows10 64bit, Version 1909 (Build 18393.900)
Nvidia Drivers Version: 442.92 Driver Studio
Daz Studio Version: 4.12.1.117 64bit
Optix Prime Acceleration: N/A
Benchmark Results
DAZ_STATS
2020-06-19 21:14:54.795 Finished Rendering
2020-06-19 21:14:54.842 Total Rendering Time: 3 minutes 56.0 seconds
IRAY_STATS
2020-06-19 21:15:25.581 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 1 (GeForce RTX 2060 SUPER): 901 iterations, 4.014s init, 228.153s render
2020-06-19 21:15:25.581 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (GeForce RTX 2060 SUPER): 899 iterations, 3.555s init, 228.424s render
Iteration Rate: [(901+899) / 228.424] 7.880 iterations per second
Loading Time: (0*3600 + 3*60 + 56.0) - 228.424 = 7.576 seconds
Note:
Device 0 used for display, and its connected on a PCI Express 2.0 x16, instead Device 1 is on a PCI Express 3.0 x16
Sorry I missed this, didn't get a notification for some reason. I ended up getting an earlier version of Daz but I've been in Blender so much recently I haven't even installed it, that E-Cycles render engine leaves iray for dead.
New Driver 451.48 release. My hardware is the same.
Windows 10 1903 Daz 4.12.1.109
CPU: i5 4690K
GPU #1: EVGA 1080ti SC2
GPU #2: MSI 1080ti Gaming <--this is my display, yes, I use GPU 2 for display.
RAM 32GB HyperX
OS Drive Samsung 860 EVO 1TB
Asset Drive: Samsung 860 EVO 1TB and WB 4TB Black HDD
2020-06-24 18:42:02.192 Total Rendering Time: 4 minutes 35.58 seconds
2020-06-24 18:42:08.709 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : Device statistics:
2020-06-24 18:42:08.709 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 0 (GeForce GTX 1080 Ti): 902 iterations, 6.272s init, 265.262s render
2020-06-24 18:42:08.709 Iray [INFO] - IRAY:RENDER :: 1.0 IRAY rend info : CUDA device 1 (GeForce GTX 1080 Ti): 898 iterations, 6.632s init, 264.731s render
Virtually identical times to the last driver, and thus still a lot slower than driver 436. I'm thinking that performance is not going to come back.
The cynic in me, is going to wonder about the upcoming release of a new GPU architecture line due soon. Pure guesswork on my part of course, although it's something we've seen before. Of course, the more an item can do, the more it's given to do.