GPU Monitoring

GPU monitoring adds a GPU Usage meter to the Dashboard so you can see how much GPU memory a host is using. It supports NVIDIA and AMD GPUs and is off by default.

Important

Install the GPU drivers on the host first, following your GPU vendor’s instructions. Official Arcane images don’t bundle vendor GPU tools: NVIDIA monitoring relies on the NVIDIA container runtime injecting nvidia-smi, and AMD monitoring reads /sys/class/drm directly.

Turn on GPU monitoring

1. Add the lines for your vendor below to the Arcane service in your compose.yaml.
2. Recreate the container.
3. Open the Dashboard and check the environment card for the GPU Usage meter.

For a remote environment, make the same changes on the agent container running on the GPU host.

NVIDIA

Expose the GPUs to the container through the NVIDIA container runtime:

services:
  arcane:
    environment:
      - NVIDIA_VISIBLE_DEVICES=all
      - GPU_MONITORING_ENABLED=true
      - GPU_TYPE=nvidia
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

AMD

Pass the DRM devices into the container:

services:
  arcane:
    environment:
      - GPU_MONITORING_ENABLED=true
      - GPU_TYPE=amd
    devices:
      - /dev/dri:/dev/dri

Check that it works

The meter shows the average memory use across detected GPUs, with the number of devices below it. When Arcane first collects stats, its logs show NVIDIA GPU detected, AMD GPU detected, or Using configured GPU type.

If the meter doesn’t appear, look for these errors in the logs:

Log message	Cause
nvidia-smi not found but GPU_TYPE set to nvidia	The NVIDIA container runtime isn’t injecting nvidia-smi. Check the NVIDIA Container Toolkit on the host.
AMD GPU not found in sysfs but GPU_TYPE set to amd	No AMD card with VRAM information was found under /sys/class/drm.
no supported GPU found	Auto-detection found no GPU tool or AMD card.

Intel GPUs aren’t supported yet. Arcane can detect one when intel_gpu_top is available in the container, but it doesn’t read Intel memory stats, so no meter appears.

Reference

Variable	Default	Description
GPU_MONITORING_ENABLED	false	Collect GPU stats.
GPU_TYPE	auto	nvidia, amd, intel, or auto. auto tries NVIDIA, then AMD, then Intel. Unknown values fall back to auto.