nvitop: An Easier GPU Monitoring & Process Management Alternative to nvidia-smi

Summary

nvitop is an interactive NVIDIA-GPU process viewer — the "one-stop solution for GPU process management". It queries device status directly through NVML Python bindings and renders a colorized terminal UI with curses, acting as a modern, richer alternative to nvidia-smi: live refresh, history graphs, process sorting/filtering, keystroke process management, a process tree view and an environment-variable screen.

• Project: https://github.com/XuehaiPan/nvitop (Apache-2.0)
• Docs: https://nvitop.readthedocs.io
• Latest PyPI release: 1.6.1

---

1. Why nvitop

nvidia-smi prints a snapshot once; tracking utilization, memory, temperature or power trends means repeatedly rerunning it or writing loop scripts. On shared-GPU AI/HPC clusters you typically need to answer three questions:

1. What is the GPU doing — utilization, memory, temperature and power of every device right now;
2. Who occupies it — which user/task runs which process on which GPU;
3. How to act — once an abnormal process is found, end it quickly.

nvidia-smi offers limited help on #2 and #3. nvitop combines all of this into a top-like interactive screen:

• Informative & fancy: shows more than nvidia-smi, colorized with box drawing;
• Monitor mode: refreshes continuously like top instead of printing once;
• Bar charts & history graphs: visualize utilization, memory, temperature and power trends;
• Process sorting & filtering: sort by utilization, memory, CPU, user and more;
• Keystroke process management: send signals to processes right in the UI;
• Process tree view: GPU processes and their parent hierarchy;
• Environment-variable & help screens;
• Mouse support.

---

2. Installation

nvitop depends on NVML (via nvidia-ml-py/pynvml) and psutil, requires Python 3.7+, and needs NVIDIA drivers installed (nvidia-smi available).

# pip install (recommended)
pip install nvitop

# upgrade
pip install --upgrade nvitop

# or via conda
conda install -c conda-forge nvitop

Then run:

nvitop

If the command is not on PATH (e.g. installed with pip3 install --user), use the module form:

python3 -m nvitop

Tip: it also starts without a GPU (CPU/memory only).

---

3. Basic Usage

Interactive monitor (default)

Run nvitop to enter a top-like screen that refreshes all devices and processes continuously.

One-shot output (like nvidia-smi)

# Print device status once (faster, lower resource usage)
nvitop -1

# Only GPU 0 and GPU 1
nvitop -1 -o 0 1

# Only devices in CUDA_VISIBLE_DEVICES
nvitop -1 -ov

# Only GPU processes with a compute context
nvitop -1 -c

Remote monitoring over SSH

ssh user@host -t nvitop

Use -t to force a pseudo-terminal and enter the interactive screen directly.

---

4. Common Key Bindings

In the interactive screen (press h for the full help screen):

Key	Action
q	Quit and return to terminal
h	Go to the help screen
a / f / c	Switch display mode: auto / full / compact
r	Force refresh the window
/	Search inside process information

Process sorting (GPU utilization, memory, CPU, user…), filtering, sending signals, the tree view and the environment-variable screen are also available.

---

5. Real-World Scenarios

Find "who took my GPU"

nvitop

Sort by memory or utilization to see processes, owners and command lines per device — no more cross-checking nvidia-smi across machines.

Cluster inspection & alerts

In monitor mode, temperature/power/utilization refresh live with history graphs, making abnormal tasks (prolonged high temperature, utilization jitter) obvious; over-threshold values are highlighted.

Quickly dispose of abnormal processes

Signal the process directly in the UI instead of opening another terminal to kill — handy on shared GPUs for fast resource release.

---

6. Advanced: Python API & Companion Tools

nvitop also ships an easy Python API for custom monitoring scripts:

from nvitop import Device

devices = Device.all()  # or Device.cuda.all() for CUDA ordering
for device in devices:
    print(device)
    print(f'  - Fan speed: {device.fan_speed()}%')
    print(f'  - Temperature: {device.temperature()}C')
    print(f'  - GPU utilization: {device.gpu_utilization()}%')
    print(f'  - Memory: {device.memory_used_human()} / {device.memory_total_human()}')
    print(f'  - Processes: {len(device.processes())}')

take_snapshots() returns all device and process snapshots in one call — handy for inspection scripts or alert integration.

The package also ships nvisel, an interactive CUDA_VISIBLE_DEVICES selector for multi-GPU environments.

---

References

• GitHub: https://github.com/XuehaiPan/nvitop
• Docs: https://nvitop.readthedocs.io
• PyPI: https://pypi.org/project/nvitop/