Overview
hw-detect inspects your machine and reports every hardware detail that affects local LLM inference: CPU capabilities, GPU topology (dedicated and integrated), unified or discrete memory, memory bandwidth, and the best available inference backend.
Example Output
Apple M4 Pro (24 GB Unified Memory):hw-detect surfaces the full GPU inventory. Integrated GPUs are visible even when the selected runtime backend resolves to CPU.
Flags
Output the full hardware detection result as a machine-readable JSON object. Useful for scripting or CI pipelines.
What It Detects
| Field | Description |
|---|---|
| CPU | Brand, core count (logical and physical), clock speed, SIMD capabilities (AVX-512, AVX2, ARM NEON) |
| GPU | Model name, VRAM, GPU type (dedicated vs. integrated), backend availability |
| Memory | Total system RAM or unified memory (Apple Silicon) |
| Backend | Best available inference backend: metal, cuda, rocm, cpu |
| Tier | Hardware tier label: ULTRA LOW, LOW, MEDIUM, HIGH, VERY HIGH, EXTREME |
| Max model size | Estimated maximum model size (GB) that fits in available memory |
GPU Topology
hw-detect enumerates GPUs into two separate lists:
- Dedicated GPUs — discrete NVIDIA, AMD, or Intel Arc GPUs with their own VRAM
- Integrated GPUs — Intel Iris Xe, AMD integrated Radeon, Apple Vega, etc.
Assist path message explains when an integrated GPU is visible but the runtime backend still falls through to CPU.
The hardware fingerprint shown at the bottom of the output is a deterministic string that uniquely identifies your hardware configuration. It is used internally for cache keying.
Supported Backends
| Backend | Hardware |
|---|---|
metal | Apple Silicon (M1/M2/M3/M4 series) |
cuda | NVIDIA GPUs with CUDA support |
rocm | AMD GPUs with ROCm support |
generic | Intel Arc and other discrete GPUs |
cpu | CPU-only inference (fallback) |