metop

A Python-based GPU/ANE monitoring tool for Apple Silicon Macs. Like nvtop or nvidia-smi, but for Metal and the Neural Engine.

Features

• GPU Monitoring (no sudo required)

  • Device, Renderer, and Tiler utilization percentage
  • GPU memory usage (in-use vs allocated)
  • Top GPU-active processes from AGX user client GPU time deltas
  • Selectable process list with keyboard and mouse-wheel scrolling
  • Real-time sparkline history

• System Monitoring

  • CPU usage and load average
  • Memory usage and swap
  • Disk usage and read/write throughput

• ANE + Power Metrics (requires sudo / powermetrics)

  • CPU/GPU/ANE/Total power (mW/W)
  • GPU/ANE frequency and active/idle residency (when available)
  • ANE utilization estimated from residency (fallback: power-based)

• Display Modes

  • Default stacked layout: top for GPU+ANE, CPU, and power; middle for memory, disk, and history; bottom for top GPU processes
  • Runtime switching with m (1 for stacked, 2 for classic)

• System Info

  • Chip detection (M1/M2/M3/M4 series)
  • CPU/GPU/ANE core counts
  • Total memory

Installation

# Install from source
pip install -e .

# Or with optional fast IOKit bindings
pip install -e ".[fast]"

Usage

# Basic monitoring (GPU only)
metop

# Enable ANE + power metrics (requires sudo)
sudo metop

# Custom refresh interval (500ms)
metop -i 500

# Disable ANE/powermetrics even with sudo
metop --no-ane

# Start in the classic layout
metop --layout classic

# Debug mode (single sample, raw output)
metop --debug

Screenshot

How It Works

GPU Monitoring

Uses IOKit via ioreg command to query the AGXAccelerator driver's PerformanceStatistics. This provides:

• Device Utilization % - Overall GPU busy percentage
• Renderer Utilization % - Shader/compute units
• Tiler Utilization % - Geometry processing
• Per-process GPU activity by diffing each AGX user client's accumulatedGPUTime

Per-process activity is reported as GPU time over the sample window, normalized into an estimated percentage. This is a best-effort approximation, not a direct hardware residency counter.

ANE + Power Metrics

Uses powermetrics (-f plist) to collect CPU/GPU/ANE power and residency/frequency data. Requires sudo because powermetrics needs root access.

ANE utilization is estimated from:

• preferred: ANE active residency reported by powermetrics
• fallback: power-based estimate

utilization = (current_power / max_power) * 100%

Requirements

• macOS Monterey (12.0) or later
• Apple Silicon Mac (M1/M2/M3/M4 series)
• Python 3.9+
• rich (terminal UI)
• psutil (memory stats)

License

MIT License