GPU Job failed

[Karol.Osinski]

Hi, today I uploaded V-Ray GPU job to render which does render locally with one RTX 3090 (runs the display too). Cloud job has failed, from the log I've read that the GPU device used for the job is Tesla T4 which is 16GB (only?).
Do you guys have any higher VRAM GPU for the cloud rendering? If yes, is it possible to somehow get a priority to access those for cloud rendering?

I had to launch this job with CPU and I'd rather not mix same project between GPU and CPU.

regards,
Karol

[vladie]

Hi Karol,

I'm sorry for the slow reply. You're correct that currently we use Tesla T4 GPU cards which have 16 GB of memory. We are working with GCP (Google Cloud Platform). When there is a better solution from GCP which we have tested and verified that it meets the Chaos Cloud's requirements, we might change it. You have identified the current solution in cases when the GPU card's memory is not enough. It is to render with the CUDA CPU Override option ON - in hybrid mode on CPU only. Just a hint - you don't need to resubmit the scene. You can clone the failed job and turn ON the CUDA CPU Override option.

Bests,
Vladie

[Karol.Osinski]

Hi vladie . I am going to try out On Demand MipMapping today. I believe CUDA CPU Override option makes the render significantly slower, doesn't it?

[vladie]

Hi Karol,

The CUDA CPU-only mode was significantly improved some time ago. So it shouldn't be a lot slower. My colleague Muhammed Hamed, V-Ray GPU Product Specialist, will soon add more details here including regarding MipMapping. Please wait till you get his advice as well.

Thanks,
Vladie

[Karol.Osinski]

Thanks vladie .

[Muhammed_Hamed]

Hey Karol.Osinski

On one hand Cuda-CPU mode uses smaller buckets now, which improves speed and avoids the buckets being stuck near the end of rendering. It should be better and faster than before
The cloud uses 16 GB GPUs at the moment, you can try the resize option if possible or On-demand mipmaps.
Let me know how this goes,

Best,
Muhammed

[Karol.Osinski]

Hi Muhammed, good to hear that. Stuck buckets can be so annoying.
I tried with On-demand mipmaps, still failed. Resize option was too much as I needed those 4K textures for the quality.

After all, I rendered it all with CPU. Took probably twice as long comparing to do it locally with RTX 3090 but I am still glad the results were almost identical to GPU.

[Muhammed_Hamed]

Did you use Cuda CPU or V-Ray renderer for that?
Cuda CPU should be identical to rendering on GPUs as you probably know

Best,
Muhammed

[Karol.Osinski]

I've used V-Ray CPU