Quick Take
Quick Take: nvitop is an interactive NVIDIA GPU monitor that replaces dense nvidia-smi tables with a live process view, compact single-line mode, and an optional Prometheus exporter. This guide covers install paths for Ubuntu and Windows, daily flags, and when nvidia-smi is still enough.
Installation
Quick Answer: Two commands are enough: pip install nvitop, then nvitop to open the monitor. The package ships on PyPI and runs on Linux, Windows and macOS wherever NVIDIA drivers are present. Ubuntu hosts that still need drivers can run the bundled install-nvidia-driver.sh script with a version number or --latest.
Quick Summary: nvitop is a Python-based interactive GPU monitoring tool built on NVIDIA Management Library (NVML). It provides an htop-like terminal interface showing real-time GPU utilization, memory usage, temperature, power draw, fan speed, and running processes across all GPUs. Unlike nvidia-smi, it supports continuous auto-refresh, process sorting/filtering, process tree views, environment variable inspection, and signal sending — all within an interactive terminal session.
pip Installation (Recommended)
Three steps take a fresh node from bare driver to a running monitor:
If Step 2 prints nothing, the driver is missing or the kernel module did not load — fix that before launching the monitor, or use the bundled script below on Ubuntu.
NVIDIA Driver Installation Script (Ubuntu)
The nvitop repository bundles a shell script for automated NVIDIA driver installation on Ubuntu systems:
Key Features
Quick Answer: Per-GPU utilization, memory, temperature, power draw, fan speed and P-state refresh continuously, with no loop script. Every GPU-occupying process is listed with PID, user, GPU memory, CPU and command line across all cards in the node, and keystrokes sort, filter, draw process trees or send signals.
1. Real-Time GPU Monitoring
Launching nvitop displays real-time metrics for every GPU in the system, including GPU utilization, memory consumption, temperature, power draw, fan speed, and performance state (P-state). All data refreshes automatically — no need to re-run commands.
2. GPU Process Viewer
nvitop shows all GPU-occupying processes with PID, username, GPU memory usage, CPU usage, resident memory, running time, and startup command. On shared GPU servers, this enables rapid identification of which user or task is consuming GPU resources.
3. Multi-GPU Support
On multi-GPU systems, nvitop displays all GPUs simultaneously with per-card utilization and memory statistics, providing unified node-level monitoring.
On shared hosts this is where the savings land. Nodes commonly carry 4x to 8x GPUs, and lab clusters routinely split one box across 3 to 8 users, so utilization numbers without ownership answers explain little. The tree view maps every slice of load to a user and PID on sight; filters isolate one tenant; a signal ends the runaway process without touching anyone else's jobs.
4. Interactive Operations
Unlike nvidia-smi which outputs static snapshots, nvitop supports keyboard-driven interaction:
- Sorting — sort by GPU memory, utilization, PID, or other columns
- Filtering — filter by specific user or process
- Process tree (t) — view GPU processes with parent-child relationships
- Environment variables (e) — inspect process environment variables
- Signal sending — send termination signals to GPU processes (with appropriate permissions)
- Help (h) — view keyboard shortcut reference
- Return (Enter) — monitor metrics for a specific process
These interactive capabilities significantly accelerate GPU resource troubleshooting.
nvitop vs nvidia-smi: Comparison
Quick Answer: Both tools display multi-GPU status at low overhead, and that is where the overlap ends. nvidia-smi needs a loop script to refresh and returns only basic process data with no interaction. nvitop refreshes itself, sorts and filters by column or user, draws a process tree, and can signal processes.
| Feature | nvidia-smi | nvitop |
|---|---|---|
| Real-time refresh | Requires loop script | Built-in auto-refresh |
| Interactive UI | No | Yes |
| GPU process details | Basic | Detailed (PID, user, CPU, mem, cmd) |
| Sorting / filtering | Not supported | Supported (by column, by user) |
| Process tree view | No | Yes |
| Multi-GPU display | Supported | Supported |
| Resource overhead | Low | Low |
| Signal sending | No | Yes |
Daily Usage Tips
Quick Answer: Plain nvitop gives the full dashboard, while nvitop -1 prints one compact line per GPU for a fast check. Adding --colorful renders utilization as spectrum bars, which reads quicker during triage. Because the same collectors import as a Python library, in-house dashboards can reuse them without shelling out.
Quick launch: Run nvitop from any terminal for an instant overview of all GPU status. Use nvitop -1 for a compact single-line-per-GPU view.
Colorful mode: Launch with nvitop --colorful for spectrum-style bar charts that visually represent utilization levels.
Python API integration: nvitop can be imported as a Python library and integrated into custom monitoring dashboards or browser-based tools, making it suitable for building in-house GPU monitoring solutions.
Conclusion
Quick Answer: nvitop replaces the nvidia-smi watch loop with an interactive, self-refreshing view of utilization, memory and per-process consumption at comparable overhead. On shared or multi-GPU nodes it shortens the path from a stalled job to the user and process responsible, which is why it belongs in the operator toolchain.
nvitop is a significant quality-of-life improvement over nvidia-smi for engineers and operators who work with NVIDIA GPUs daily. Its interactive interface, real-time refresh, process management capabilities, and low overhead make it an essential tool for GPU infrastructure monitoring. Whether you manage a single workstation or a multi-GPU server cluster, adding nvitop to your toolchain provides immediate visibility into GPU resource utilization and helps accelerate problem diagnosis.
Key Takeaway
Key Takeaway: Install nvitop with pip and run it to replace nvidia-smi with an interactive view of GPU, memory and per-process usage. Three keys carry most of the debugging value: h for help, t for the process tree, e for environment variables. BENQUAN GLOBAL LIMITED supplies NVIDIA GPU servers with competitive pricing.
RFQ checklist before you PO
Quick Answer: Freeze model and option SKUs, support or subscription years, ship-by date, and HA or spare needs on one dated ATP before you compare channel prints. Use the checklist below, validate lines on SKU Lookup when the brand is indexed, then open a BOM risk RFQ — soft channel prices stay industry estimate or pending confirmation.
- Model / MPN lines — chassis and options matching this page (not a vague family name).
- Quantities + HA — single unit vs pair; spare FRUs if required.
- Support / subscription years — tier and term on the same quote as hardware.
- Ship-by / quote validity — lock reprice-until-ship language when memory or list cliffs apply.
- Acceptance — Service Tag / serials vs PO; reject surprise third-party parts without paperwork.
Frequently Asked Questions
Quick Answer: Installation is a single pip command, and the payoff over nvidia-smi is an auto-refreshing htop-style view of utilization, memory, temperature, power draw, fan speed and P-state, plus the PID, user and command behind every GPU process. Because it wraps NVML, the same collectors can feed a custom dashboard.
Where is the official nvitop GitHub and how do I install it?
The official repo is github.com/XuehaiPan/nvitop. Install with pip install nvitop, then run nvitop. This page is the operator path — pip, the Ubuntu driver script, and nvitop versus nvidia-smi — so you launch before you compare. Confirm nvidia-smi prints a GPU table first.
What does nvitop offer over nvidia-smi?
Built-in auto-refresh instead of loop scripts, an interactive htop-like UI, detailed per-process data (PID, user, GPU memory, CPU, command), sorting and filtering by column or user, a process tree view, and signal sending — at comparably low resource overhead.
Which metrics does nvitop display in real time?
Per-GPU utilization, memory consumption, temperature, power draw, fan speed and performance state (P-state), all refreshing automatically, plus every GPU-occupying process with PID, username, GPU memory usage, CPU usage, resident memory, running time and startup command.
Can nvitop feed custom monitoring dashboards?
Yes. nvitop is built on the NVIDIA Management Library (NVML) and can be imported as a Python library, so its collectors integrate into custom monitoring dashboards or browser-based tools. For quick visual triage, launch with the colorful flag for spectrum-style utilization bars.