Iray - complex scenes - I don't know why I bother
krickerd
Posts: 188
Just can't use a large set else defaults to CPU. I got 3, count em 3 1080s. That's 24GB video RAM. I run reduce texture size. Claims it reduced 101 images. Still not using the cards. I don't know why I bother.
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
The snag is you don't have 24GB of VRAM available for rendering. You have 8GB of VRAM available which, assuming your scene fitted into it, would be rendered considerably faster than one 1080. VRAM isn't numerically summed to produce a larger number unless you have a set of cards which support NVLINK which, you guessed it, the 1080 doesn't. One of those 1080s will lose at least 1GB of that to Windows, possibly more, limiting all of the others and, as they're not RTX cards, the RTX emulation code will be loaded onto them losing you roughly 2GB more. I'd be surprised if you had more than 5GB free for your renders.
You effectively have 5GB of VRAM which really isn't enough these days unless you're willing to sacrifice a lot of detail or use older models. It's not that you *can't* get something good from an 8GB GTX card - you can. It's just becoming increasingly more difficult.
I would consider a 12GB RTX card to be entry level these days.
Just check you DS log, you'll know why... VRAM consumption is allocated to all cards, while the Display Card takes heaviest burden - i.e. Windows + App + Viewport + allocated VRAM from the scene + 2GB consumption from iray Engine itself...
If you run out VRAM on you Display Card, GPU based render will be failed, then it falls back to CPU.
Install GPU-Z to monitor Memory Used from time to time, and you probably have to heavily optimize your scene.
you can try to render scene in parts (to be stitched together in post) and also hide anything out of frame that is not needed as a shadow caster/light emitter. Some additional restrictions are that you will need to preserve overlapping content across the seams of the renders you stitch together, and also preserve any major shadow casters/light emitters for all renders, so it can be quite strategic process. Getting all of that to fit on 8GB card can also be hard depending on scene requirements. Even a single character can be more than your limitations, depending on character, outfit, subdivision.
Sometimes it pays to render out character at low resolution for whole scene pass, and then do HD+hair etc as a single additional pass only for the visible parts of the body (to be composited in post). Can also hide any parts of character covered by clothing.
Again, working with limitation of card requires some strategic decision-making. Or export to blender where you wont have VRAM limitation (i think?). Or get card with more VRAM.
What we haven't done is offer a solution. Here's mine. Sell two of your 1080s and buy an RTX 3060 12GB card or an RTX 4070 12GB / RTX 4080 16GB if your budget will allow. Allocate your remaining 1080 to Windows duty and use your new RTX card for rendering as a compute device only. You will be surprised at the difference this makes to what can be rendered on the GPU. 1080 cards are relics as far as iRay is concerned. Keep one, buy an RTX. The days of 3 x GTX cards being useful is long since over I'm afraid.
well i offered up the solution of compositing scene from partial renders or using blender as renderer.
I only have a 6GB RTX and just use compositing as my friend - something that Iray excells at. It renders very clean alpha!
Okay show me how to render this: https://www.daz3d.com/flooded-cemetery without running out of VRAM. Ran reduce texture script on all of it. Didn't matter.
Wait, so the GPU power is being used on all cards but VRAM isn't????!!!! And nope, aint upgrading unless forced. My 3 card system is custom built and water cooled. Though 6.5 years old, I'm squeezing every year I can out of it.
nearly all my animated videos use png alpha image series of characters
not always compiled im a video editor, unlike DAZ studio, Carrara, iClone and other softwares (including Blender) support animated textures so I use billboards rendered in DAZ studio iray
(I leave the ground shadow off if going to be a billboard)
a lot of my scenes are parallax too
similar to what Dartanbeck did but I actually place the rendered background, mid and foreground planes of buildings, tree etc in the scene
Cards are modified by me. Custom water cooled as stated in other reply. I think I'd be better off using Vue for backgrounds since I know going in the render is CPU, then using Iray for characters in FG when appropriate. When the cards are being used, they are pretty fast since all 3 are in fact used. I've verified that with Riva Tuner / Afterburner and the system temp goes up to about 35 to 37C. Sometimes I still use 3Delight when I can't be bothered to convert old mats especially since generic conversion settings are rarely very close.
That looked pretty fluid. That was made with hundreds of PNG stills? That's impressive. If I did more animation I think I would try filament. Haven't really tried it yet. It's fast.
3 image series (yes thousands all up) and an AI generated BG made into a 3D plaque using a depth map
https://convert.leiapix.com/#/local/1 + Zbrush Alpha to mesh
the depth map thing doable with renders too but I use Carrara or Octane render as iray canvases never worked for me
an old video I just found on depthmap parallax
now we can use AI to do depthmaps
from 3:35 still relevent
Scene-based VRAM consumption from this product is less than 1.4GB, so the allocated consumption on each of your 3 cards is around 0.45GB. You don't need to optimize it. Monitor the VRAM Used on you display card with GPU-Z before you click render. Share your DS log.
VRAM of all cards is used, but VRAM is not pooled unless you use Nvlink. Each card loads the scene in its own VRAM.
Iray renders light paths on the GPU (if the scene fits) or on the CPU (if seelcted or if falback allowed) and sends the result back to the main process, which integrates the successive samples to build up the final render. Adding more cards allows it to process more samples concurrently, reducing render time, but each is a fully independent process (indeed, if there are cards with different effective memory capacities then one may drop out while the others continue). So both memory and power are used, neither is actually pooled.
Geforce 10 Series is Nvlink 1.0 compatible however Nvlink 1.0 is not mature. Even if you get a link, you can just link 2 cards, and there're still some memory pooling bugs in the driver...
I'm getting smaller memory usage reported!
This is on Windows 11 and a 3060.
With viewport set to 'texture preview' I get around 5300MB, starting an HD render (from that texture preview) is 6750MB. That might just fit on a 1080, making the render size even smaller is a possible way to keep the GPUs rendering.
Just out of interest, I squeezed the viewport to its smallest size and set it to Iray, that reached a similar 6770MB on the GPU. Launching an HD render from that small Iray preview, the GPU memory requirement was 7950MB.
(I also set a 4k Iray viewport, which reached 7700MB, LOL)
Another method to keep VRAM use to a minimum is queueing jobs on the server, closing Studio and starting the queue without Studio running.
The same Flooded scene at 1920x1080 from my previous post now uses only 5000MB!
Unless you have the cards running a physical NVLink, they won't do memory pooling. You have 8GB of VRAM (in reality, more like 6,5. The rest is reserved by Windows).
8GB cards are practically useless for larger scenes, unless you're ready to make the extra effort and render the scene in parts, as someone else mentioned.
I remember that my first GPU for rendering was a 1080TI. The 11GB ram it had was BARELY enough for a nice looking render.
Now I have the 4090, and trust me, 24GB is not a large amount either. By today's standards, and 4K, 8K, 16K textures becoming the norm, it's time to make peace with the fact that gaming cards are not suited for stuff like DAZ, Blender, etc.
Sell your cards while they have any value left. And buy a used RTX3090, they go for less than $700 sometimes (and, the 3090 will still be approx 30% faster than your 3 GPU's combined).
A system with a single air-cooled RTX 4090 will not only allow you to render larger scenes but will also leave anything else in the dust, speed-wise.
It is nice that you are trying to hang on to your three GTX 1080 cards, but those cards are at the very tail of what is currently supported in NVIDIA CUDA (which is used for Iray). NVIDIA could decide to drop support for Maxwell and Pascal architectures in the next major CUDA release (both are already EOL), and then you are just an update away from DAZ Studio no longer working on your cards at all.
It would be better to sell those three cards while you can still get some value from them and invest in a single better card which has more memory and RTX.