Maniphest T101041

Error: Launch in CUDA queue copy_from_device (integrator_intersect_closest)
Closed, Archived
Actions

Assigned To
None
Authored By
Luke A. Guest (Lucretia)
Sep 13 2022, 7:21 PM
Tags
  • BF Blender
Subscribers
Luke A. Guest (Lucretia)
Richard Antalik (ISS)

Description

System Information
Operating system: Windows-10-10.0.19044-SP0 64 Bits
Graphics card: Quadro K2100M/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 425.91

Blender Version
Broken: version: 3.3.0, branch: master, commit date: 2022-09-06 15:39, hash: rB0759f671ce1f
Worked: (newest version of Blender that worked as expected)

Short description of error
With the Blender 3.3.0 splash .blend file, on rendering with CUDA GPU I get the following error, nothing is displayed on screen.

Error: Launch in CUDA queue copy_from_device (integrator_intersect_closest)

This was using nvidia driver 425.91.

I also tested with 3.3.1, same issue.

With the 3.4 beta, it displayed the start of an image and then failed with the same error.

I have since upgraded to a newer driver, 426.78, and that fails with a different error, That's nothing rendered and error:

Error: Launch failed in CUDA queue copy_from_device (integrator_shade_surface_raytrace integrator_sorted_paths_array integrator_compact_shadow_states prefix_sum)

Exact steps for others to reproduce the error

  1. Load the blender_3_3_lts_splash_by_piotr_krynski.blend file.
  2. Set settings to CUDA GPU
  3. F12

Related Objects

Event Timeline

Luke A. Guest (Lucretia) created this task.Sep 13 2022, 7:21 PM
Richard Antalik (ISS) changed the task status from Needs Triage to Needs Information from User.Sep 14 2022, 12:43 AM
Richard Antalik (ISS) added a subscriber: Richard Antalik (ISS).

Checking used resources, this file used 4GB of GPU RAM, while your GPU seems to support 2GB RAM at maximum. Please check if your GPU is running out of memory.

Richard Antalik (ISS) merged a task: T101042: Launch failed in CUDA queue copy_from_device (integrator_shade_surface_raytrace integrator_sorted_paths_array integrator_compact_shadow_states prefix_sum).Sep 14 2022, 12:44 AM
Luke A. Guest (Lucretia) added a comment.Sep 14 2022, 7:57 AM

The K2100M only has 2GB RAM.

Richard Antalik (ISS) closed this task as Archived.Sep 15 2022, 3:35 AM

Thanks for info. Will close since this file requires more resources than available. Even though these messages could be improved on, currently we do not considering to be a bug. Thanks for reporting though.

Luke A. Guest (Lucretia) added a comment.Sep 15 2022, 8:18 AM

I don't know anything about CUDA, I do know about swdev though, but is there no way to manage this via some sort of memory management to stop it using all the memory in one go?

Richard Antalik (ISS) added a comment.Sep 16 2022, 2:44 AM

I am not very knowledgable when it comes to GPU compute technologies. I am aware, that cuda and optix can share memory. Not sure if this has to be managed by engine and to what extent. Looking at cycles docs, I see only option to share memory between 2 GPUs with NVLink.

I should have also mention, that error messages like this can be caused if GPU is overclocked, so you may check that as well. However I would probably start with simplified file, and increasing complexity while checking memory usage to be fairly sure whether this happens truly due to available memory or it is some tangentialy related issue.

Luke A. Guest (Lucretia) added a comment.Sep 16 2022, 11:22 AM

It's a stock laptop, not overclocked.

I don't know how to "start with simplified file" in this instance. Other things render fine, the bmw, the white lands file, I didn't leave lone monk running long enough to test it fully.

Richard Antalik (ISS) added a comment.Sep 16 2022, 11:47 PM

Just disabling objects from rendering should do the trick, that is non destructive way.