otel-amdgpu-metrics

A Rust library that collects AMD GPU metrics and exports them through OpenTelemetry.

Introduction

This project provides OpenTelemetry instrumentation for AMD GPUs on Linux. It reads metrics directly from the sysfs filesystem (/sys/class/drm/) so there's no need to install ROCm or any other external dependencies. If your system has the amdgpu kernel driver loaded, this should work.

I built this because GPU metrics are often left out of observability setups or require vendor-specific tooling. With this library you can get GPU telemetry into whatever backend you're already using with OpenTelemetry.

This is NOT an official OpenTelemetry project. Just a community contribution.

Metrics

The following metrics are collected for each detected AMD GPU:

Metric	Description	Unit
hw.gpu.utilization	GPU core utilization	%
hw.gpu.memory.used	VRAM currently in use	bytes
hw.gpu.temperature	GPU temperature	Celsius
hw.gpu.power	GPU power consumption	watts

Each metric includes an hw.id attribute with the card identifier (e.g., card0, card1).

Requirements

• Linux with the amdgpu kernel driver
• Rust 1.85 or later (uses 2024 edition)

Tested on EndeavourOS with a Radeon RX 7900 GRE (discrete) and Ryzen 7 9700X integrated graphics.

Usage

Add to your Cargo.toml:

[dependencies]
otel-amdgpu-metrics = "0.1"
opentelemetry = "0.31"

Then initialize in your code:

use opentelemetry::global;
use otel_amdgpu_metrics::init;

let meter = global::meter("my-app");
let gpus = init(&meter).expect("Failed to detect AMD GPUs");

println!("Monitoring {} GPU(s)", gpus.len());

The library will automatically detect all AMD GPUs using the amdgpu driver and register metrics with your OpenTelemetry meter provider.

Running the Example

cargo run --example basic

License

Apache-2.0