1080ti GPU too old for ML on new immich versions?
I'm running into ML failures with my GTX 1080 Ti (Pascal architecture). Getting constant CUDNN_STATUS_EXECUTION_FAILED errors from the machine learning container. Looking at the immich Dockerfile, there is an explicit: "
Even after downgrading to NVIDIA driver 535 (CUDA 12.2) which fully supports Pascal, the containerized CUDNN 9.x still fails
- Error occurs in cudnnReduceTensor operations during ML inference
What I've Tried:
- Downgraded from driver 570 → 535 (CUDA 12.8 → 12.2)
- Driver and host CUDA support Pascal fine
- Container's CUDNN 9.x doesn't support Pascal regardless??
For context: This used to work fine before system updates. I recently updated to Ubuntu 24.10 and to Immich current release
From asking Claude about the commit history:
System Info:
Is this an issue I should report, or am I just versioned out and should use CPU for ML tasks? Thanks!
RUN apt-get update && \
** # Pascal support was dropped in 9.11**
apt-get install --no-install-recommends -yqq libcudnn9-cuda-12=9.10.2.21-1 && \
apt-get clean && \
rm -rf /var/lib/apt/lists/*
Even after downgrading to NVIDIA driver 535 (CUDA 12.2) which fully supports Pascal, the containerized CUDNN 9.x still fails
- Error occurs in cudnnReduceTensor operations during ML inference
What I've Tried:
- Downgraded from driver 570 → 535 (CUDA 12.8 → 12.2)
- Driver and host CUDA support Pascal fine
- Container's CUDNN 9.x doesn't support Pascal regardless??
For context: This used to work fine before system updates. I recently updated to Ubuntu 24.10 and to Immich current release
From asking Claude about the commit history:
September 11, 2024 (commit ad58d7e23) - They downgraded from CUDA 12.3 to 12.2, This is when the line first appeared (without version pinning), because the 12.2 runtime image didn't include cudnn9.
October 20, 2025 (commit bcfdb2f9d) - They pinned the version to prevent automatic upgrades:
# Pascal support was dropped in 9.11
apt-get install --no-install-recommends -yqq libcudnn9-cuda-12=9.10.2.21-1 && \
So the line has existed for about 13 months, but was only version-pinned 8 days ago.
System Info:
- GPU: NVIDIA GeForce GTX 1080 Ti (11GB VRAM, Pascal architecture, compute capability 6.1)
- OS: Ubuntu 24.10 (Oracular Oriole)
- Current Driver: 535.247.01 (CUDA 12.2)
- Previous Driver: 570.133.07 (CUDA 12.8) - came with Ubuntu upgrade
- Immich Version: release (latest)
Is this an issue I should report, or am I just versioned out and should use CPU for ML tasks? Thanks!
