RTX speed increase?

[Morbid Angel]

Eh well, every single project I do, I at the end open the scene and try on gpu, 99% it crashes, or does not work, or if it does takes just as long to render.

[Joelaff]

Last time i checked, hybrid was slower than everything else....but to be honest hybrid mode isnt the original question, dont make me do another test pleassee

Sorry. You don't need to do a test if you don't want to.

We have experienced the exact opposite with VRay5, where hybrid is by far the fastest option. I guess it is all very hardware dependent.

[Joelaff]

I'm not sure how some one can use gpu in production anyway. The limitations and instability quickly outweight any speed gains *(which are marginal) over cpu.

For our current project, which involves a lot of underwater stuff, the GPU is a LOT faster (3-5x), and even just using CUDA on the CPUs is marginally faster.

But, yes, many headache with regard to features that don't work, or more with regard to the mix of GPUs on the render farm. That is the real issue for us. The workstations with modern cards actually work pretty well.

It's all very scene dependent. I do think GPU will eventually be, as Howard Hughes said, "the way of the future." Right now it is very bleeding edge, and there is blood all over the offices here.

[squintnic]

RTX always slower for me using CUDA w/ 2990wx

GPU rendering is faster in some scenes (good on product stuff) atm i mostly use it to render passes like smoke and volumes which are slow on CPU - cos it just dies on any complex VFX or arch vis shot with more than a few maps. I wont even go into load times for maps and the OOC implementation...

I hope in future it gets some more attention.

Perhaps it needs to be its own product ? I know no one would buy it right now if it was released as a stand alone renderer. The only reason we are using it is cos its bundled with CPU. Anyone doing serious GPU rendering is in Redshift

[Joelaff]

That's what we mainly use it for as well. Volumetric stuff. Right now for an underwater scene with millions of bubbles (stuff splashing into the water form above).

I think tiled/MIP Mapped textures like .tx or tiled EXR should help a lot with the texture loading issues, especially if you bucket render. But we really NEED independent user control over bucket sizes for GPU, with independent control for the GPU bucket size and the CPU bucket size. The default bucket sizes are pretty absurd, and not having control over that is just silly. (Says this control freak.)

[Morbid Angel]

there is blood all over the offices here.

I had a good laugh well I suppose for certain things like you said volumetric stuff it works. But I will quickly find a limitation in any area of the gpu. For example, I had to render space craft flying through kilometres of clouds. The clouds were pretty high res voxel count individual clouds out of houdini. I had maybe 5-6 different clouds which I then instanced around. Each of the clouds took about 5Gb of ram on cpu, so 6*5=30Gb for just clouds. But here is a catch, I had to have the space craft act as a holdout to produce volumetric depth for comp. And the space craft it self was also 40-50 GB of ram. Made ouf of hundreds of alembic caches, thousands of subdivision meshes, combined render with 3d motion blur was over 90Gb of ram. Need I say more?

[Joelaff]

Yeah, the memory is a big limitation. Of course, if you work smart you can squeeze a lot more out of it, but working "smart" in this way is not always efficient for time or money. It's usually cheaper just to stuff more RAM in your boxen (or in your boxers, though the folks at the computer stores seem to frown on that approach) than to try to work around lack of it.

[Morbid Angel]

Joelaff, this is precisely what I did. I went on ebay and ordered 4 TB of DDR3 ram for the farm, upgrading every node to 128GB. It cost me around $3000. Though slower, but as you know not by much I still rely heavily on cpu. I'd rather get the proj rendered in 24 hours, then fight with gpu for the issues it has and maybe 2-3 times speed up it offered. This is a big one for me, gpu was suppose to promise x10 or whatever speed up, but its not really true at all

[Morbid Angel]

And while we were on this subject I just happened to have a gpu setup scene open so I did a couple of tests. Here is the same scene (simple scene with 3 trees) rendered using BF/LC default settings. I've lowered cpu vray settings to be 1/8 0.01 and got a fairly clean image in 3m 1sec I have dual xeon e5-2683v3 24 core 48 thread. I then rendered the scene with sample limit of 500 on two cards I have RTX 2080ti and GTX 1080 ti and got render time of 1m 53 seconds. Then I rendered the same scene with 250 sample limit on just the RTX which produces a bit noisy result but kind of matches that of cpu and got 1m 56 sec. Granted I'm rendering with vray 4.3 (I must stick with this one for current project) But you can see exactly what's being said above, rtx is faster but not nearly as fast as one would expect. And certainly rendering a clean image with two cards at 66% of cpu render time really isn't worth the struggle at the moment for me anyway. THIS IS STRICTLY IMHO.

[glorybound]

I just tested on my 8 minutes interior GPU/RTX scene. 18 Minutes with hybrid/CUDA rendering.

[glorybound]

I can work fine in GPU, but eventually and on every job, it'll crash and I won't be able to use GPU anymore. I think the latest error is a 700 error.

[Karol.Osinski]

Calling VRay GPU "production ready" has been triggering me since the first time they annouced it at Total Chaos 2018...Production ready...but WHAT kind of production? I am sorry to say it but definitely not complex architectural visualization which is the biggest Chaos Group customer base by the way. It is still so unstable if you work on heavy scenes and I tried it plenty of times.

Just the fact it often takes AGES to load IPR or Production render in CUDA / RTX drives me crazy...

[Karol.Osinski]

I then rendered the scene with sample limit of 500 on two cards I have RTX 2080ti and GTX 1080 ti and got render time of 1m 53 seconds.

Wait, is there Sample Limit parameter in VRay for Maya?

[Muhammed_Hamed]

Wait, is there Sample Limit parameter in VRay for Maya?

I wrote about that on Facebook few days ago, there is no samples limit in Maya.. but you can do the conversion manually

[Karol.Osinski]

Thank you Muhammed_Hamed ! I use fb less and less these days I had to miss this one. Will definitely look into it!