Nvidia 555 drivers kills CUDA

Support Technical Support
1

I can’t see this having been mentioned anywhere. I use Arch linux (BTW :wink: ), and have asked on the forums there, but nada. However, I’m thinking a Blender problem.

Nvidia drivers 555 (doesn’t matter the sub variant) are failing with CUDA errors. This on Blender 4.2.

Illegal address in CUDA queue copy_from_device (integrator_init_from_bake integrator_shade_surface integrator_sorted_paths_array)

Reverting to 550 drivers and all is good.

Except it gets weirder. Blender 4.1 seg faults.

# Python backtrace
# Blender 4.1.1, Commit date: 2024-04-15 15:11, Hash e1743a0317bc
Read library:  '/home/steve/.config/blender/4.2/scripts/addons/botaniq_full/blends_280/flowers/Flower_Daisy_A_spring-summer.blend', '//../../.config/blender/4.2/scripts/addons/botaniq_full/blends_280/flowers/Flower_Daisy_A_spring-summer.blend', parent '<direct>'  # Info
Read library:  '/home/steve/.config/blender/4.2/scripts/addons/botaniq_full/blends_280/grass/Grass_Tall_A_spring-summer.blend', '//../../.config/blender/4.2/scripts/addons/botaniq_full/blends_280/grass/Grass_Tall_A_spring-summer.blend', parent '<direct>'  # Info
Read library:  '/home/steve/.config/blender/4.2/scripts/addons/botaniq_full/blends_280/Library_Botaniq_Materials.blend', '//../../.config/blender/4.2/scripts/addons/botaniq_full/blends_280/Library_Botaniq_Materials.blend', parent '/home/steve/.config/blender/4.2/scripts/addons/botaniq_full/blends_280/grass/Grass_Tall_A_spring-summer.blend'  # Info
Read library:  '/home/steve/.config/blender/4.2/scripts/addons/botaniq_full/blends_280/weed/Weed_Dandellion_A_spring-summer.blend', '//../../.config/blender/4.2/scripts/addons/botaniq_full/blends_280/weed/Weed_Dandellion_A_spring-summer.blend', parent '<direct>'  # Info

# backtrace
blender(+0x207290b) [0x616a9198790b]
blender(+0x15fe239) [0x616a90f13239]
/usr/lib/libc.so.6(+0x3cae0) [0x7a4e48f65ae0]
/usr/lib/liboslexec.so.1.13(_ZN9OSL_v1_133pvt9LLVM_Util13call_functionEPN4llvm5ValueEN16OpenImageIO_v2_54spanIKS4_Lln1EEE+0x41) [0x7a4e5339b441]
/usr/lib/liboslexec.so.1.13(+0x1490be) [0x7a4e533490be]
/usr/lib/liboslexec.so.1.13(+0x15505f) [0x7a4e5335505f]
/usr/lib/liboslexec.so.1.13(+0x22e69a) [0x7a4e5342e69a]
/usr/lib/liboslexec.so.1.13(+0x15d58a) [0x7a4e5335d58a]
/usr/lib/liboslexec.so.1.13(+0x22e69a) [0x7a4e5342e69a]
/usr/lib/liboslexec.so.1.13(+0x17f369) [0x7a4e5337f369]
/usr/lib/liboslexec.so.1.13(+0x18a476) [0x7a4e5338a476]
/usr/lib/liboslexec.so.1.13(+0x6ae3b) [0x7a4e5326ae3b]
/usr/lib/liboslexec.so.1.13(+0x6d05c) [0x7a4e5326d05c]
/usr/lib/libstdc++.so.6(+0xe0c84) [0x7a4e492e0c84]
/usr/lib/libc.so.6(+0x92ded) [0x7a4e48fbbded]
/usr/lib/libc.so.6(+0x1160dc) [0x7a4e4903f0dc]

# Python backtrace

Even with 550 drivers.

I can’t share a scene - uses Botaniq, so has paid for assets.

Tried re-installing CUDA and no change.

For now, I’ve pinned nvidia, nvidia-utils (and not that it matters, lib32-nvidia-utils) at 550, but any ideas?

The reason I’m thinking a Blender fault, is that everything else tested works just fine. Cyberpunk 2077, the three most recent unigine benchmarks (heaven, valley and superposition).

2

It must be something to do with Linux, as I run the 556.12 driver and it’s just fine in Windows…

3

Given the eerie silence both here and on the Arch forums, I suspect it’s a “me” problem somewhere.

4

When was the last time you didn’t have this problem, and what has changed since?

5

Honestly, it’s only with the 555 drivers. 550 drivers work just fine. There has been no update to CUDA, so it’s something with the driver.

At least, with Blender 4.2. I don’t often use the stable version, so can’t say when that broke, but it’s a different error.

6

I assume you have been to the NVidia Boards already…but it might be a good idea to post it if you haven’t already…

7

Running the 555.99 Studio drivers here, but under Windows. Still for what it’s worth, not had any issues.