Skip to content

15-privileged-linux-amd.md

Latest commit

 

History

History
80 lines (61 loc) · 3.24 KB

15-privileged-linux-amd.md

File metadata and controls

80 lines (61 loc) · 3.24 KB
Raw

Privileged Metrics — Linux + AMD

GPU, power, and thermal on a Linux host with an AMD GPU — discrete Radeon, or the Strix Halo iGPU (Ryzen AI Max, e.g. an HP ZBook Ultra G1a). For the shared helper + systemd deployment, see Privileged Metrics (overview); this guide is the AMD specifics.

What you get, and from where

Metric Source Needs root?
gpu.util, gpu.vram, gpu.temp, power.gpu amd-smi if installed, else amdgpu sysfs no
npu.util (XDNA) — (no stable counter yet) —
power.cpu RAPL via the helper yes
temp.pkg hwmon (k10temp / zenpower) via the helper yes

Both GPU sources are unprivileged, so you get the GPU panel out of the box — installing amd-smi only adds richness and cross-version stability.

Two sources, tried in order

  1. amd-smi (preferred) — AMD's tool, shipped with ROCm. The daemon runs:

    amd-smi metric --usage --power --temperature --mem-usage --csv

    Columns are matched by token (e.g. gfx_activity, socket_power, edge_temperature, used_vram/total_vram), so the parser survives the header renames that differ across amd-smi releases.

  2. amdgpu sysfs (automatic fallback) — the in-tree amdgpu driver always exposes these, no package required. Fills any field amd-smi did not provide, or the whole set when amd-smi is absent:

    sysfs node Metric
    …/card*/device/gpu_busy_percent gpu.util
    …/device/mem_info_vram_used + …_total gpu.vram
    …/device/hwmon/hwmon*/power1_average (µW) power.gpu
    …/device/hwmon/hwmon*/temp1_input (m°C) gpu.temp
    …/device/hwmon/hwmon*/freq1_input (Hz) gpu.clock
    …/device/hwmon/hwmon*/pwm1 (0–255 duty) gpu.fan (as a percent)

    The daemon picks the first DRM card whose device/uevent says DRIVER=amdgpu. gpu.fan is absent on APUs with no dedicated GPU fan (shared cooling) — it shows up on discrete Radeon cards that expose pwm1.

Install amd-smi

# Arch / CachyOS
sudo pacman -S rocm-smi-lib                     # provides amd-smi
# Debian / Ubuntu (after adding the AMD ROCm apt repo)
sudo apt install amd-smi-lib
# Fedora
sudo dnf install amd-smi

amd-smi version                                  # confirm it is on PATH
./bin/heimdall-daemon --once | grep -E "gpu|power|npu"
# e.g. gpu.util=37%  gpu.vram=25%  gpu.temp=48C  power.gpu=12.5W  npu.util=unavailable

amd-smi needs no root for these read-only queries — run the daemon as your normal user. The helper is only for the CPU extras (RAPL power.cpu, hwmon temp.pkg).

Unified memory caveat

On a Strix Halo iGPU, mem_info_vram_* reflects the GPU's GTT/VRAM carve-out, not system RAM. Heimdall reports it as gpu.vram (percent of that budget) and keeps system memory on the separate mem.* keys — the two are never conflated.

NPU (XDNA / Ryzen AI)

The XDNA NPU (amdxdna driver) has no stable utilisation counter yet, so npu.util reads unavailable on AMD hosts — the accelerator is advertised but never faked. It will light up when the driver exposes a residency figure (the same situation as the Apple ANE npu.util gap).