Hardware acceleration support

Immich can offload video transcoding and machine learning to a GPU. The two use different backends and have different constraints, listed separately below.

Setup instructions are in Set up hardware transcoding and Set up hardware-accelerated machine learning.

Transcoding

Supported APIs

• NVENC (NVIDIA)
• Quick Sync (Intel)
• RKMPP (Rockchip)
• VAAPI (AMD / NVIDIA / Intel)

Limitations

• The instructions and configurations here are specific to Docker Compose. Other container engines may require different configuration.
• Only Linux and Windows (through WSL2) servers are supported.
• WSL2 does not support Quick Sync.
• Raspberry Pi is currently not supported.
• Two-pass mode is only supported for NVENC. Other APIs will ignore this setting.
• By default, only encoding is currently hardware accelerated. This means the CPU is still used for software decoding and tone-mapping.
  • You can benefit from end-to-end acceleration by enabling hardware decoding in the video transcoding settings.
• Hardware dependent
  • Codec support varies, but H.264 and HEVC are usually supported.
    • Notably, NVIDIA and AMD GPUs do not support VP9 encoding.
  • Newer devices tend to have higher transcoding quality.

Machine learning

Supported backends

• ARM NN (Mali)
• CUDA (NVIDIA GPUs with compute capability 5.2 or higher)
• ROCm (AMD GPUs)
• OpenVINO (Intel GPUs such as Iris Xe and Arc)
• RKNN (Rockchip)

Limitations

• The instructions and configurations here are specific to Docker Compose. Other container engines may require different configuration.
• Only Linux and Windows (through WSL2) servers are supported.
• ARM NN is only supported on devices with Mali GPUs. Other Arm devices are not supported.
• Some models may not be compatible with certain backends. CUDA is the most reliable.
• Search latency isn't improved by ARM NN due to model compatibility issues preventing its use. However, smart search jobs do make use of ARM NN.