OTOY Forums

Hi everyone,

I just want to add something important in Out Of Core (OOC) feature and riser 1X : speed are very affected !
Results on a scene with 7-8GB of textures:
- with one Titan on the motherboard (16x or 8x) : 6.5Ms/s (DL)
- with one Titan Black on amfeltec (4X -> 4 x 1X) : 3.05 Ms/s (DL)
- with one 780Ti on amfeltec : 0.78Ms/s (DL) (seems that's theres too many texture to load)
- with one 780Ti on USB3 riser : 0.78Ms/s during 5s and then freeze the OS.

Note : I have also edited registry to increase graphic time-out : "TdrLevel" and "TdrDelay" as some card on the 1x often failed during rendering.

Thanks for sharing

itou31 wrote:Hi everyone,

I just whant to add something important in Out Of Core feature and riser 1X : speed are very affected !
Results on a scene with 7-8GB of textures:
- with one Titan on the motherboard (16x or 8x) : 6.5Ms/s (DL)
- with one Titan Black on amfeltec (4X -> 4 x 1X) : 3.05 Ms/s (DL)
- with one 780Ti on amfeltec : 0.78Ms/s (DL) (seems that's theres too many texture to load)
- with one 780Ti on USB3 riser : 0.78Ms/s during 5s and then freeze the OS.

Note : I have also edited registry to increase graphic time-out : "TdrLevel" and "TdrDelay" as some card on the 1x often failed during rendering.

Hello itou,
This is strange... once the scene is loaded there should be minimal I/O activity since most of the processing happens on the GPU. Here's my theory... 7-8Gb is a lot of textures - more than any of your cards could natively handle. Out of core memory kicks in, changing the I/O activity. The Titans with 6Gb can handle the majority of your textures. The Ti's can't.

I have not looked into how out of textures work in detail but perhaps one of the devs could chime in with a bit of advice?

Seeker

Yes, I have a big scene and wanted to try OOC. With the plugin OcDs or standalone (exported from OcDs) textures are very big on character model (many in 4k and even some are 8k).
The trick is to rescale texture as needed (often 1/4). I will try to reduce to at least under 6GB to fit to Titans, but difficult to reduce to under 3GB.
I think that the texture are already "compressed" into VRAM ?

itou31, are the cards you have all 6GB? Forgive me if I am stating the completely obvious, but If you have just one that is 3GB then you are capped at 3GB across the board. You mentioned something about 3GB, that's why I am saying...
I am not aware what OOC Memory does, haven't tried turning it on/off or messing with settings...

Hi Notiusweb,

I have 3 titans on x8 motherboard (6GB)
and 2 titan black on amfeltec extender (PCIe4X to 4x PCie1X) so 6GB also
and 3 780Ti (2 on the amfeltec) and one on the USB3 riser (3GB).

If you want to try, take a "big" scene with many texture that could not fit on 6GB (your titanZ) and just try to render.
It should fail (notice in the system log tab).
To enable OOC on the OcDS, just move the slider out of core to reseve RAM, move also the headroom slider enough to move at leat half of texture to the RAM system. If I understand, the headroom is reserved the VRAM for other thing. So when OcDs fill the VRAM of data, i will stop at "max VRAM minus headroom" and then let the rest in the system RAM. You will see on the graph only after activating wiewport to allow OcDs to compile data.
I don't know if there is a button only for compiling the scene before the rendering, just to check the VRAM status, because, if the render begin, it will fail all graphic drivers and cannot restart the drivers (should then reboot the PC).

Hi itou31,
Just a naive thought...Did you ever try rendering a larger scene while not using the 3GB cards? In other words, did you ever try de-selecting the 3GB cards from the active CUDA rendering devices in order to see if you could get a more workable flow at 6GB Vram, be it with less CUDA cores? Wondering how your experience would be straight with 6GB...

Hi,

yes I have already test all scenarios.
I thought that what I described are clear.

Results on a scene with 7-8GB of textures:
- with one Titan on the motherboard (16x or 8x) : 6.5Ms/s (DL)
- with one Titan Black on amfeltec (4X -> 4 x 1X) : 3.05 Ms/s (DL)
- with one 780Ti on amfeltec : 0.78Ms/s (DL) (seems that's theres too many texture to load)
- with one 780Ti on USB3 riser : 0.78Ms/s during 5s and then freeze the OS.

Each result is one shot with only one board selected. Restart DAZ or Octane and fresh reload the scene. I note that select and deselect often in the same session would crashed DAZ/OcDs.
I also test with only 3 titan, only 2 titan black, all titan together.
If 780Ti together : crash and freeze (due to the 780ti on USB riser).
Same ratio speed results.

So If you can try (need at least 4 DAZ characters with cloth to fill the 6GB VRAM), and report if you have the same issue with your riser. Perhaps there are good and bad quality riser (I don't think so) ?

I see, okay. You are running just one card at a a time trials. I thought maybe it was mixed somehow.
Then I will really look to see if I get something like this, as I am working with 6GB Vram.

- with one Titan on the motherboard (16x or 8x) : 6.5Ms/s (DL)
- with one Titan Black on amfeltec (4X -> 4 x 1X) : 3.05 Ms/s (DL)

However, it will actually be one 12GB at 16x (Titan X) vs one 6GB at 1x (Titan Z) on the riser...but anyway, just to see how it goes with the OOC, I'll let you know what I see, then we can take it from there with my results.

Okay, I completed my testing. I don't have a Z on motherboard to run vs a Z on riser, so I just used my X to see how it would go down with a high memory scene.
And you know what I found? Even with the X, I had a hard assembling a 6GB memory scene. I used 5 G3F characters, a background, and 4 of 5 had clothes. So I guess one could say, well add a 6th then to push it over 6 GB, test with a Z, and see how it works with the OOC. But Daz gets so damn slow, even with the X and its 12 GB memory, as I closed in on 6GB, that it wasn't even worth it for me at this point. I can't even navigate the app, let alone the OcDS plugin at that level. I did run a test, though, and this is what I recorded. Don't know if it serves any purpose, don't think it ever used OOC.


But I can totally imagine that a 3GB card would break down under the weight, and add that it's on a riser at 1x, it would crack. And from what I am seeing, OOC is not going to be as optimized as we would like.
But, I mean, forget it....It's so slow it's unusable at the 6GB level, and my scene didn't even need OOC. And even if I could mirror your results with a Z on motherboard vs Z on Amfeltec, what does that do, just confirms what I already know. It's slow as F^@% either way...
Hey, how do you navigate a scene on Daz at that memory level, wireframe? Shutting off stuff in the scene editor? (ie - hair)...let me know if there is a way to speed things up in the app itself without sacrificing texture, polygon count etc.

To any non-DAZ users, long story short, not only are you battling a riser, but a app and respective OR plugin that is super slow at 5-6 GB texture levels. In fact, probably would be too slow with a 32 GB Pascal...Well, maybe not, but for me with the X, too slow no matter what at this level.

But here's my silver lining,I did learn something. I did export the scene from DAZ as an .obj mesh and then loaded that mesh into the OR standalone, and then I did not have to deal with the DAZ app's crawl. Felt much smoother to work with. I thinkl the conversion compressed something, looked like it was less memory...Here is my render at same res, higher speed than the plugin. I don't know if that extra process matters to you, but render time was faster 0:57 vs 3:14, same 2560 x 1440 resolution, tried to mirror all lighting elements.
Sorry to be such a Daz grouch. Okay, here is the finished image (done in standalone)
BTW, I just loaded it in sloppy, for testing purposes...nah, just kidding, I worked super hard on every detail so that I would 'Tantalize' you...: