GPU Monitoring

Turn on GPU monitoring to add a GPU Usage meter to the Dashboard showing a host’s GPU memory use. 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 your vendor’s lines below to the Arcane service in compose.yaml. NVIDIA exposes GPUs through the NVIDIA container runtime; AMD passes in the DRM devices. For a remote environment, change the agent container on the GPU host instead.
2. Recreate the container.
3. Check the Dashboard environment card for the GPU Usage meter.

NVIDIA

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

AMD

services:
  arcane:
    environment:
      - ARCANE_GPU_MONITORING_ENABLED=true
      - ARCANE_GPU_TYPE=amd
    devices:
      - /dev/dri:/dev/dri

Check that it works

The meter shows average memory use across detected GPUs, with the device count below. On first collection, the logs show NVIDIA GPU detected, AMD GPU detected, or Using configured GPU type. If the meter doesn’t appear, look for these errors:

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.

Reference

Variable	Default	Description
ARCANE_GPU_MONITORING_ENABLED	false	Collect GPU stats.
ARCANE_GPU_TYPE	auto	nvidia, amd, intel, or auto. auto tries NVIDIA, then AMD, then Intel. Unknown values fall back to auto. Intel isn’t supported yet: Arcane detects it via intel_gpu_top but reads no memory stats, so no meter appears.