Enable Distribute memory across device Denoise Optix Error #119959

Closed
opened 2024-03-27 13:21:48 +01:00 by abraham castilla · 7 comments

System Information
Operating system: Windows-10-10.0.19041-SP0 64 Bits
Graphics card: NVIDIA GeForce RTX 2080 Ti/PCIe/SSE2 NVIDIA Corporation 4.6.0 NVIDIA 551.86

Blender Version
Broken: version: 4.1.0, branch: blender-v4.1-release, commit date: 2024-03-25 20:42, hash: 40a5e739e270
Worked: (newest version of Blender that worked as expected)

Short description of error
Enabling "Distribute memory between devices" in the Optix selection system option and the "Render>Denoise> Optix" option, gives the error "Failed to enable peer access on CUDA context (Peer access already enabled)"

Exact steps for others to reproduce the error

  • with 2x 2080 TI RTX graphics cards linked with NVIDIA Link Bridge.
  1. Enable Preferences>System> CUDA or Optix>"Distribute memory between devices".
    2 .Enable "Render>Cycles>GPU" and "Render>Sampling>Denoise>Optix"
**System Information** Operating system: Windows-10-10.0.19041-SP0 64 Bits Graphics card: NVIDIA GeForce RTX 2080 Ti/PCIe/SSE2 NVIDIA Corporation 4.6.0 NVIDIA 551.86 **Blender Version** Broken: version: 4.1.0, branch: blender-v4.1-release, commit date: 2024-03-25 20:42, hash: `40a5e739e270` Worked: (newest version of Blender that worked as expected) **Short description of error** Enabling "Distribute memory between devices" in the Optix selection system option and the "Render>Denoise> Optix" option, gives the error "Failed to enable peer access on CUDA context (Peer access already enabled)" **Exact steps for others to reproduce the error** - with 2x 2080 TI RTX graphics cards linked with NVIDIA Link Bridge. 1. Enable Preferences>System> CUDA or Optix>"Distribute memory between devices". 2 .Enable "Render>Cycles>GPU" and "Render>Sampling>Denoise>Optix"
abraham castilla added the
Type
Report
Status
Needs Triage
Severity
Normal
labels 2024-03-27 13:21:48 +01:00
Philipp Oeser added the
Interest
Render & Cycles
label 2024-03-27 15:28:53 +01:00
Member

Hi, thanks for the report. Unable to confirm with single NVIDIA GPU.
Could you confirm this when single GPU is connected or selected in preferences?
Wait, I misunderstood something. Option peer_memory/Distribute memory between devices is not available for me 🙂 (looks like multiple GPUs are required for this option to show up in prefs)

**System Information**
Operating system: Windows-10-10.0.22631-SP0 64 Bits
Graphics card: NVIDIA GeForce RTX 3050 Laptop GPU/PCIe/SSE2 NVIDIA Corporation 4.6.0 NVIDIA 551.76
Hi, thanks for the report. Unable to confirm with single NVIDIA GPU. ~~Could you confirm this when single GPU is connected or selected in preferences?~~ Wait, I misunderstood something. Option `peer_memory/Distribute memory between devices` is not available for me 🙂 (looks like multiple GPUs are required for this option to show up in prefs) ``` **System Information** Operating system: Windows-10-10.0.22631-SP0 64 Bits Graphics card: NVIDIA GeForce RTX 3050 Laptop GPU/PCIe/SSE2 NVIDIA Corporation 4.6.0 NVIDIA 551.76 ```

Correct, this happens with two cards connected with the "Bridge" of "Nvidia Link". In my case two RTX 2080 TI. The "Distribute memory between devices" option appears when you have them connected.

Correct, this happens with two cards connected with the "Bridge" of "Nvidia Link". In my case two RTX 2080 TI. The "Distribute memory between devices" option appears when you have them connected.

the latest revision with Blender 4.1.1 "810bfcc30a93" from April 2, the problem persists when activating "render>denoise>Opendenoise>GPU activate" and then changing this option to Optix.

the latest revision with Blender 4.1.1 "810bfcc30a93" from April 2, the problem persists when activating "render>denoise>Opendenoise>GPU activate" and then changing this option to Optix.
Member

Thanks. I'll ask/find someone to verify this on multi-gpu setup.

Thanks. I'll ask/find someone to verify this on multi-gpu setup.
Member

My guess is this issue was introduced when Cycles switched over to using the "primary CUDA context" 4f58cffb4e.

CC @pmoursnv do you have time to take a look at this?

My guess is this issue was introduced when Cycles switched over to using the "primary CUDA context" 4f58cffb4efd0ba574fe8f6d8905cc406ce22be0. CC @pmoursnv do you have time to take a look at this?
Member

Patrick found the fix !120255 , confirming.

Patrick found the fix !120255 , confirming.
Blender Bot added
Status
Resolved
and removed
Status
Confirmed
labels 2024-04-15 12:19:01 +02:00
Member

@ThomasDinges hi, can we still backport fixes to 4.1.1? This one looks useful.

@ThomasDinges hi, can we still backport fixes to 4.1.1? This one looks useful.
Sign in to join this conversation.
3 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: blender/blender#119959
No description provided.