Skip to main content

NVIDIA

NVIDIA GPUs with NVENC move video encoding off the CPU and onto dedicated silicon. If you have a GTX 1050 or newer, you've got NVENC. The encode is not the phase that dominates a run, though — see Encoding quality for what the card actually buys you.

What you get​

  • NVENC encoding: h264_nvenc, hevc_nvenc. Offloads encoding to dedicated hardware on the GPU.
  • NVDEC decoding: hardware-accelerated decode, keeps the full pipeline on GPU.
  • CUDA scaling: scale_cuda resizes frames on the GPU instead of pulling them back to CPU.
  • CUDA scene analysis: when OpenCV has CUDA support and hardware.gpu_analysis is on, frame differencing for scene detection runs on the GPU. Face detection stays on the CPU (OpenCV Haar cascades) — there is no CUDA face path.
  • Taichi title rendering: with the gpu extra installed, Taichi picks the CUDA backend (Vulkan second) for animated title screens. This is the phase that costs the most on a CPU-only box.

Requirements​

  • NVIDIA GPU (GTX 1050+ / any RTX)
  • CUDA drivers installed
  • FFmpeg built with NVENC support (most distro packages include this)

Check if everything's working:

immich-memories hardware

If NVENC is available, you'll see it listed with the specific encoders found.

Configuration​

hardware:
  enabled: true
  encoder_preset: "balanced"   # fast | balanced | quality
  gpu_decode: true
  gpu_analysis: true

Nothing to select: NVIDIA is probed first, so if NVENC works it is used. On a multi-GPU host pick the card with CUDA_VISIBLE_DEVICES / NVIDIA_VISIBLE_DEVICES — there is no device_index in the config.

In Docker​

The image does not bundle drivers; the NVIDIA Container Toolkit injects them. NVENC needs the video capability on top of the default compute,utility:

services:
  immich-memories:
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu, video]
    environment:
      NVIDIA_DRIVER_CAPABILITIES: compute,video,utility

Then docker compose exec immich-memories immich-memories hardware should list h264_nvenc. See Linux + NVIDIA for a full compose file.

Encoding quality​

NVENC quality is slightly below software libx264 at the same bitrate, but for memory videos the difference is invisible, so take the speed.

Just don't buy the card for the encode. Encoding is the smaller half of a CPU-only run: title rendering was ~263 s of a ~339 s assembly at --cpus=2, and analysis was 7.4 of 10.1 minutes end to end. The bigger wins from this GPU are Taichi title rendering and CUDA scene analysis. See CPU-Only Mode for the measured split.