- This topic has 28 replies, 10 voices, and was last updated 3 years, 9 months ago by
.
-
Posts
-
This topic came up on Discord. Should someone who has a GTX 1060 and is upgrading to a higher end RTX 4000 or RTX 3000 card keep their old GTX 1060 in their system to improve Daz rendering speed?
Some posts on the Daz forums and elsewhere with other people testing a benchmark scene show that keeping the mid to low end GTX 1000 series card after adding a RTX 3000 series card actually hinders render speed.
I don't have a GTX 1060 or a RTX 4070, but I have a 3090 and a GTX 1050 ti. I decided to try it. This was tested in Daz 4.21.0.5. I ran the single 3090 test first. Then I booted into safe mode, ran DDU to remove the drivers, rebooted and reinstalled the latest Nvidia drivers. Rebooted into windows, confirmed windows saw the cards, rebooted again, then ran the same Daz rendering benchmark scene.
Iteration Rate: (DEVICE_ITERATION_COUNT / DEVICE_RENDER_TIME) iterations per second
Single 3090: (1800/112.118) = 16.054 iterations per second
3090 and 1050 ti:
3090: (1681/120.251) = 13.979 iterations per second
1050 ti: (119/116.275) = 1.023 iterations per secondUsing a GTX 1050 ti with a RTX 3090 reduces the 3090's Iray render speed(iterations per second)
Also, if you use two RTX 3090s at the same GPU clock speed and same VRAM speed, you don't get twice the performance of a single 3090. There is some overhead when you use multiple cards.
Daz render log info can be found in: C:\Users\YORU USER NAME\AppData\Roaming\DAZ 3D\Studio4
I doubt its a driver issue. Its probably a Daz/Iray issue.
Also, as said above, there is some overhead when you use multiple cards. In Daz with Iray, two identical cards don't give you twice the render speed of a single card. Therefore, two cards that aren't similar won't give you the same speed of adding up the render rate of each card individually.
I am absolutely not surprised and I would never use two GPUs at the same time if they have different performances. There are actually two problems, but both are related to the fact that to properly render 3D, you must have information from all the potential light sources, that is the complete scene, as any reflection on any object can have an impact on the rendering of any point. All these information is stored in the internal memory of every GPU (VRAM).
The first problem related to that is that the GPU with the smallest VRAM will limit the available RAM of the whole scene. If you have a GPU with 8Gb and a GPU with 24 Gb that work together to render a scene, you will not be able to render a scene that requires more than 8Gb. This may not be a problem for some scenes, but this is frequently a serious limitation.
The second problem is that the state (texture) of all object must be up to date to render properly a scene. If a GPU has found that an object is now, say, reddish because of some lighting/reflection, the other GPU must be informed. Hence after a rendering iteration (or several depending on the way the software is written), the different GPU must exchange information to update their current view of the scene. This exchange is limited to the modified textures (if the program is well written), but it requires that the GPUs stop working, transfer memory segments between them and restart. This can be a significant slow down that is not compensated by the extra computing power if the GPUs have very different speed. For the same reason, rendering with the GPU and the CPU is rarely a good idea, except if your CPU is extremely fast and that this compensates the GPU lost cycles due to the data update. If your GPU are identical, the global increase of computing speed makes this operation interesting, but the requirement of data exchange limits the speed-up that is always <2.
Thought that there was a thread (or threads) somewhere within the forums (here or DAZ, or maybe even other) about this at some point. Seemed like something was talked about at this level and regarding question of both hardware and drivers, but that it landed somewhere with possibly SLI being deficient since a certain time? Or not compatible or updated enough for RTX? Can't remember where seen.
(Sorry for vagueness. Trying to jog others memories since mine is failing on the topic.)The RTX 4000 series consumer cards don't support SLI/NVLink. They don't have the hardware for it. The RTX 3090 and RTX 3090 TI support SLI/NVLink in a limited number of games and benchmarks.
Some software can use SLI/NVLink to pool memory between cards.
Interesting stuff. One question though, is it possible to use the secondary just for the RAM and pool the 2 RAM's into one, and disable the GPU on the secondary. At least you could benefit from fitting a bigger scene using the main card without worrying about maxing out RAM and dropping to CPU render mode.
@Gts6 Memory pooling with GPUs requires the 2 GPUs to be linked via SLI (NVLink), and 2 identical graphics cards (same model and series). Personally I'd just use the 1050 for the monitor and the 3090 for DS.
@frank22, Yeah... My 2070 super has 8b vram. When I render in Daz, I have to restart first whenever I have a scene that takes up more than 4gbs worth or else it drops to cpu. Makes me regret not getting a cpu with a built in gpu. So scene optimizer is the way I'm goin. The new 4000 series cards doesn't really appeal at the price tag combined with the meagre vram increases on the low and mid tier cards. Was tempted to get a 3060 with 12gbs but the 2070 is faster.
I'm pretty much in the same situation with an 8GB 2080. Some recent scenes rendered fine with GPU for the first one, then goes to CPU (even in Iray preview) after that.... so restart DS and it works OK.
This is such a regular occurrence that I don't think it's just because the scene's marginally not fitting to VRAM, but that DS has a serious memory leak, and/or the GPU memory is not flushing between renders. The memory purge script does absolutely nothing, it's supposed to flush VRAM but doesn't. It could actually be an Iray (NVIDIA) issue, who knows but it's fucking annoying anyway.Was hoping Windows 11 would reserve less vram than 10. Nope.
I don't suppose Daz could make use of that magic whereby allocating normal ram to virtual vram.
@littlepleasures, is the gpu memory flushing efficiently on your end?I don't know why people talk about SLI here. As someone else already stated, SLI has NOTHING to do with CUDA based multi GPU workloads. The only common thing is... both tech use multi GPUs. That stops right there.
NVLink could be interesting because it does support CUDA workloads... and the VRAM pooling thing is cool... but yeah, you have to have the right hardware.
I always had two and even at one time three GPUs while working with DAZ and Iray. Same gen GPUs, or one gen difference. I started with two GTX 970 (effing crap but it got me started). Then two 980Ti. Then two 1080Ti (with for a while a 1080 as a 3rd card but performance didn't improved much so I sold it). Then a Titan RTX and a 1080Ti. And finally as of now a 3090 with a Titan RTX. Each time I did tests, each time the double GPU setup won by a long shot.
I just did a test with a scene...
With Both GPUs: 3m50s
Only the 3090: 7m01s
Only the Titan RTX: 7m21sThat shows two things: 1) it's very close to double the speed and 2) that Titan RTX was a beast for its time, seeing how close it is to the huuuuuuge 2.5 slots 3090!!!
And BTW the Titan RTX runs at PCIe 3.0 4X and the 3090 at PCIe 4.0 16X. That also shows how bus speeds don't influence render times that much, even when loading up large scenes. And yes I also did multiple tests when I built my rig, before I decided to set it up this way.
Now... there's certainly an overhead to multi GPUs setups, and if your 2nd GPU can't even compensate for it and provide some more horses... it's worthless... and could probably in fact be a nuisance. Although I never experienced it myself except maybe for that time when I had 3 GPUs and maybe then the overhead was too high to be worth it.
whenever I have a scene that takes up more than 4gbs worth or else it drops to cpu. Makes me regret not getting a cpu with a built in gpu.
A CPU with a built in GPU wouldn't make a difference because the built in GPU doesn't support Iray. There doesn't seem to be any significant difference in CPU rendering when you use the same CPU with and without a built in GPU.
I always had two and even at one time three GPUs while working with DAZ and Iray. Same gen GPUs, or one gen difference.
I found that some versions of Daz are buggy when you get to three and four GPUs, even GPUs of the same manufacturer and model. Sometimes Daz won't even detect the 4th GPU. Yet other 3D rendering software on the same systems detects all four GPUs without issue.
And BTW the Titan RTX runs at PCIe 3.0 4X and the 3090 at PCIe 3.0 16X. That also shows how bus speeds don’t influence render times that much, even when loading up large scenes. And yes I also did multiple tests when I built my rig, before I decided to set it up this way.
PCIE Generation and PCIE 8X vs PCIE 16X doesn't matter in Daz Iray because all of the data is loaded to the VRAM before the render starts. The only thing I found that has a noticeable impact in render times is GPU core overclocking. GPU VRAM overclocking has a minimal impact in render times.
PCIE Generation and PCIE 8X vs PCIE 16X doesn’t matter in Daz Iray because all of the data is loaded to the VRAM before the render starts.
I agree. Even 4X doesn't change much, if anything. I never tested 2X though.
Nice to learn from everybody. Thanks guys, saves a lot of time chasing a rabbit down the hole.
My thinking on having a built-in gpu on the cpu being advantageous is that if using it for display only, it would free up some of the vram that Windows 10 is reserving.
I read that Windows 10 reserves a considerable amount of vram of each card and that Windows 7 didn't have that issue.
Hope I'm not derailing the thread too much. I would love to know why the vram isn't flushing as it should on one setup but not on another.
After deleting an asset that overloaded the vram and being back to what had worked it goes to cpu. Waiting and trying again doesn't help.
Assuming the Daz versions are the same, either settings are wrong somewhere or hardware isn't as compatible?
My thinking on having a built-in gpu on the cpu being advantageous is that if using it for display only, it would free up some of the vram that Windows 10 is reserving.
I would love to know why the vram isn’t flushing as it should on one setup but not on another.
Reported VRAM usage in Windows and most software such as MIS Afterburner is often misleading. Windows allows program to reserve more VRAM than is actually in use. You will see windows and other software report a certain amount of VRAM to be in use when its actually available for use by other software and is just "reserved" until its needed elsewhere.
There are some games and other programs that will reserve huge quantities of VRAM when its not needed. However if you run other software, the VRAM suddenly becomes available.
Also, I've had Windows report large amounts of VRAM usage and other software report that Windows is only using 0.7GB of VRAM.
- You must be logged in to reply to this topic.
