# System Requirements ## Supported Platforms **Operating Systems** - Ubuntu 22.04 LTS or newer - Debian 12 or newer - RHEL 9 or newer - Any Linux distribution with kernel 5.4+ The shipped Docker build targets **Linux/amd64**. Telemetry uses Linux-specific paths (`/sys`, `/proc`, `/dev/dri`). Running Docker inside a VM on another operating system exposes the VM's environment, not the physical host's Windows or macOS telemetry. The capabilities endpoint (`GET /api/v1/capabilities`) is available wherever the service runs, but hardware-enriched values depend on Linux host visibility (`/sys`, `/proc`, and optionally `/dev/dri`). **Hardware Platforms** - Any x86-64 processor (Intel or AMD) - ARM64 images/packages are not supplied by the current build: the Telegraf archive and Debian package target are amd64. - Intel Arc GPU (optional, for GPU metrics via qmassa) - Intel NPU (MTL/ARL/LNL/PTL generations, optional, for NPU metrics) ## Minimum Requirements | Component | Minimum | Recommended | | -------------- | ---------------- | --------------------- | | Processor | 2 cores @ 1 GHz | 4 cores @ 2 GHz | | Memory (RAM) | 512 MB | 2 GB | | Disk Space | 2 GB (for build) | 10 GB (with headroom) | | Docker | 24.0+ | 26.0+ | | Docker Compose | 2.20+ | 2.25+ | ## Software Requirements **Required Software** - Docker 24.0 or newer (`docker --version`) - Docker Compose 2.20 or newer (`docker compose version`) - `git` for cloning the repository - `curl` for testing endpoints (optional but recommended) **Optional Software** - `make` — for convenience commands (e.g., `make helm-lint`) - `kubectl` — if deploying to Kubernetes - `helm` 3.8+ — if using the Helm chart for Kubernetes deployment ## Hardware-Specific Notes ### OpenVINO inference in the extended image The extended image supplies OpenVINO plus Intel OpenCL/Level Zero GPU and NPU userspace runtimes and compilers. The default image collects GPU/NPU telemetry without these inference dependencies. - GPU inference requires a supported Intel GPU, its host `i915`/`xe` driver, and access to `/dev/dri/renderD*`. - NPU inference additionally requires compatible host `intel_vpu` and firmware, and access to `/dev/accel/accel*`. PMT telemetry alone does not prove inference access. The image does not install firmware or kernel modules. - The shipped privileged Compose profile exposes the devices. A restricted deployment must explicitly grant the required device nodes and render-group permissions; do not change host security settings merely to pass a test. - Official Intel vendor binaries target Ubuntu, but their pinned dependencies are resolved against Debian 13 without adding Ubuntu repositories or replacing the system C library. This is project-level compatibility validation, not an upstream support statement for Debian. - Real CPU/GPU/NPU inference was validated on an Intel Core Ultra 5 225H with Arrow Lake graphics/NPU and Linux 7.0.0-42. Other device, firmware and kernel combinations require their own validation. See the [hardware smoke test](./testing.md#openvino-accelerator-smoke-test). Missing devices fail that test explicitly; it does not fall back to CPU. ### Intel Arc GPU (Recommended for GPU metrics) - Requires Intel Arc GPU to be installed on the system - Requires a compatible host kernel driver (`i915` or `xe`) and DRM access. - qmassa collects device data; `mm-plugin-gpu` parses its FIFO stream into telemetry. Power and temperature can additionally require perf/MSR access. - The optional extended image adds userspace OpenVINO/VAAPI components; it does not install or replace the host kernel driver. **Collected metrics:** - Engine usage (compute, render, copy, video, video-enhance) - GPU frequency - GPU power consumption **Capabilities output note:** - GPU entries are also exposed in capabilities JSON (`minimal` and `expanded` profiles). - PCI branding/model naming uses `lspci` (`pciutils`) when available. - `lspci` does not require `sudo` for read-only inventory. If GPU is absent, qmassa logs `No DRM devices found` and exits gracefully. Other metrics continue normally. ### Intel NPU (Optional, for NPU telemetry) - Supported on Meteor Lake (MTL), Arrow Lake (ARL/ARL-H/ARL-S), Lunar Lake (LNL), and Panther Lake (PTL) - Requires Intel NPU driver (`intel_vpu`) to be loaded on the host - Requires `/sys/class/intel_pmt/` to be accessible inside the container (provided by `privileged: true` and `/sys:/sys:ro` in `compose.yaml`) **Collected metrics:** - NPU power draw (watts) - NPU frequency (Hz) - NPU temperature (degrees Celsius) - NPU utilization (%), when supported - NPU bandwidth (bytes/second) - Tile configuration - Memory usage (bytes), when the required telemetry fields are present. Unsupported readings are omitted; measured zero remains zero. The native collector does not publish `-1` as a substitute for unavailable memory. See the [native metric migration guide](./native-metric-migration.md) for field names and display conversions. If NPU is absent or driver is not loaded, the NPU reader logs a warning and enters idle mode. Other metrics continue normally. ## Capabilities API Quick Check After deployment, verify platform/device capabilities with: ```bash curl -s http://localhost:9090/api/v1/capabilities?profile=minimal | jq curl -s http://localhost:9090/api/v1/capabilities?profile=expanded | jq ``` Use `minimal` for a compact platform/device summary and `expanded` for full technical inventory. ## Validation Checklist - [ ] Linux kernel 5.4+: `uname -r` - [ ] Docker installed: `docker --version` - [ ] Docker Compose installed: `docker compose version` - [ ] At least 512 MB RAM available: `free -h` - [ ] At least 2 GB disk space: `df -h /` - [ ] (Optional) Intel GPU: `lspci | grep -i intel | grep -i graphics` - [ ] (Optional) Intel NPU driver: `ls /sys/bus/pci/drivers/intel_vpu/` If all checks pass, you are ready to proceed with the [Get Started Guide](../get-started.md). ## License Copyright (C) 2025-2026 Intel Corporation SPDX-License-Identifier: Apache-2.0