dgxtop

繁體中文 | 简体中文 | 日本語

High-performance, interactive system monitor for NVIDIA DGX systems with real-time GPU, CPU, memory, disk, and network monitoring.

dgxtop is a comprehensive system monitoring tool purpose-built for NVIDIA DGX infrastructure. It provides real-time visibility into GPU utilization, VRAM, temperature, power draw, NVLink topology, and system resources — all in an interactive terminal UI. Built in Rust with direct NVML access for maximum performance and reliability.

Quick Install

curl -fsSL https://raw.githubusercontent.com/DennySORA/dgxtop/main/install.sh | bash

See Installation for more options.

Why dgxtop?

• Direct NVML Access — Reads GPU metrics through NVIDIA Management Library, not nvidia-smi subprocess calls. Faster, more reliable, and more detailed.
• DGX-Optimized — Supports multi-GPU monitoring, NVLink topology, ECC error tracking, and PCIe throughput — features critical for DGX A100/H100/B200 and DGX Spark.
• Full System View — CPU per-core utilization, memory (RAM + Swap), disk I/O (IOPS, latency, throughput), and network interfaces in a single dashboard.
• Interactive Process Management — Sort, filter, and kill GPU processes directly from the TUI. See per-process GPU utilization, VRAM usage, and host memory.
• Secure by Design — No subprocess shell-outs, PID recycling protection, config value sanitization, and UTF-8 safe rendering. Passed deep security audit.

Features

GPU Monitoring (via NVML)

Category	Metrics
Utilization	GPU %, Memory controller %, per-process SM utilization
Memory	VRAM used/total/free, BAR1 usage, per-process GPU memory
Bandwidth	Memory bandwidth utilization (actual/theoretical GB/s), unified-memory auto-detection (LPDDR5X), PCIe TX/RX throughput with 1/6/12/24h avg/max stats
Thermal	Temperature (with slowdown/shutdown thresholds), fan speed, 1/6/12/24h avg/max
Power	Draw/limit (watts), usage %, total energy consumption (kWh), 1/6/12/24h avg/max/cumulative energy
Clock	Graphics, SM, Memory, Video frequencies (current/max MHz)
State	Performance state (P0–P15), throttle reasons, compute mode, persistence mode
Health	ECC errors (corrected/uncorrected), retired pages (SBE/DBE)
Topology	NVLink active links with remote GPU mapping, PCIe Gen/Width
Codec	Encoder/Decoder utilization %
Identity	UUID, serial number

System Monitoring

Category	Metrics
CPU	Aggregate and per-core usage (htop-style), user/system/iowait breakdown, temperature, frequency, load average (1/5/15m), tasks (running/total)
Memory	RAM used/total, buffers, cached, available, swap usage
Disk I/O	Per-device read/write throughput, IOPS, await latency, sorted by throughput
Network	Per-interface RX/TX throughput, errors, sorted by activity

Time-Series Analytics

• Line charts for GPU utilization, memory, temperature, and power with Braille markers
• 1/6/12/24h statistics — avg/max for CPU, Memory, GPU utilization, temperature, power (with cumulative kWh), memory, PCIe TX/RX; avg R/W for Disk; cumulative ↓/↑ for Network
• Progressive display — stats windows only appear after enough data has been collected
• Minute-resolution aggregation — memory-efficient 24h storage (~28KB per metric)

Interactive TUI

• Three Views — Overview dashboard, GPU detail (full-page per-GPU), full-screen process table
• Vim Keybindings — Navigate with j/k, switch tabs with 1/2/3, GPU selection with h/l
• Device Selection — Cycle network interfaces (n/N), disk devices (d/D) with chart/stats following selection
• Process Management — Sort by GPU mem/utilization/CPU/PID, filter by name, kill with confirmation
• Responsive Layout — Auto-adapts to terminal size, conditionally shows charts/stats based on available space
• Visual Design — Rounded panels, gradient gauges, line charts, alternating row colors, color-coded thresholds

Multi-GPU Overview with Active Workloads

Installation

Quick Install (Recommended)

curl -fsSL https://raw.githubusercontent.com/DennySORA/dgxtop/main/install.sh | bash

The installer auto-detects your libc and picks the matching release target. For NVIDIA GPU metrics, glibc (-gnu) builds are recommended.

Download Binary

Download pre-built binaries from GitHub Releases:

Platform	Architecture	Download
Linux	x86_64 (glibc, recommended)	dgxtop-x86_64-unknown-linux-gnu.tar.gz
Linux	x86_64 (musl, compatibility)	dgxtop-x86_64-unknown-linux-musl.tar.gz
Linux	ARM64 (glibc, recommended)	dgxtop-aarch64-unknown-linux-gnu.tar.gz
Linux	ARM64 (musl, compatibility)	dgxtop-aarch64-unknown-linux-musl.tar.gz

Note: On some systems, musl builds may fail to load NVIDIA NVML (libnvidia-ml.so), resulting in missing GPU metrics.

Build from Source

git clone https://github.com/DennySORA/dgxtop.git
cd dgxtop
cargo build --release
# Binary: target/release/dgxtop

Cargo Install

cargo install --git https://github.com/DennySORA/dgxtop.git

Usage

Basic Usage

# Start with default settings
dgxtop

# Custom refresh interval (0.5 seconds)
dgxtop -i 0.5

# Disable GPU monitoring (system metrics only)
dgxtop --no-gpu

# Use green color theme
dgxtop -t green

Command Line Options

Option	Description	Default
-i, --interval <SECS>	Update interval in seconds (0.1–10.0)	1.0
-t, --theme <NAME>	Color theme: cyan, green, amber	cyan
--no-gpu	Disable GPU monitoring	false
--net-max <N>	Max visible network interfaces / disk devices (1–20)	3
--log-level <LEVEL>	Log level: error, warn, info, debug	warn

Keyboard Shortcuts

Key	Action
q / Ctrl+C	Quit
Tab / Shift+Tab	Switch between views
1 / 2 / 3	Jump to Overview / GPU Detail / Processes
j/k or ↑/↓	Navigate up/down
h/l or ←/→	Select GPU (in GPU Detail view)
s	Enter sort mode
r	Reverse sort order (in sort mode)
/	Filter processes by name/PID/user
K	Kill selected process (with confirmation)
e	Toggle per-core CPU view (htop-style)
n / N	Cycle network interface
d / D	Cycle disk device
+ / -	Increase / decrease refresh speed
?	Toggle help overlay

Views

Overview — Responsive dashboard with CPU gauge (load average, tasks, htop per-core toggle), memory bars, GPU cards, disk I/O and network panels with line charts and 1/6/12/24h statistics. Device selection highlights chart/stats for the chosen interface or disk.

GPU Detail — Full-page single-GPU view with comprehensive metrics: utilization, VRAM, memory bandwidth (actual/theoretical GB/s), BAR1, thermal (with thresholds), power (with energy kWh), throttle reasons, P-state, all clock domains (Graphics/SM/Memory/Video), PCIe info, NVLink topology, encoder/decoder, ECC, retired pages, compute mode, UUID — plus time-series charts and per-GPU process list.

Processes — Full-screen GPU process table with sortable columns, search filter, and process kill capability. Shows PID, user, GPU device, type, GPU%, VRAM, CPU%, host memory, and command.

Architecture

┌─────────────────────────────────────────────────────────────────────┐
│                           Main Thread                               │
│  ┌───────────┐    ┌────────────┐    ┌────────────────────────────┐ │
│  │  AppState  │◄───│  UI Loop   │◄───│ crossbeam channel (rx)    │ │
│  │            │    │  (ratatui) │    │ bounded(256)               │ │
│  └───────────┘    └────────────┘    └────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────────┘
                                                ▲
                     Tick events + Key/Mouse    │
                                                │
┌───────────────────────────────────────────────┴─────────────────────┐
│                        Event Thread                                  │
│  ┌───────────────────────────────────────────────────────────────┐  │
│  │  crossterm::event::poll + Tick timer → AppEvent → channel tx  │  │
│  └───────────────────────────────────────────────────────────────┘  │
└─────────────────────────────────────────────────────────────────────┘

Collectors (called on each Tick):
  ├── GpuCollector        (NVML: utilization, temp, power, clocks, memory, bandwidth,
  │                         ECC, PCIe, NVLink, BAR1, throttle, P-state, encoder/decoder,
  │                         retired pages, energy, UUID, serial)
  ├── GpuProcessCollector (NVML + /proc: per-process GPU/CPU/memory stats)
  ├── CpuCollector        (/proc/stat + /proc/loadavg: per-core usage, frequency,
  │                         temperature, load average, task count)
  ├── MemoryCollector     (/proc/meminfo: RAM, swap, buffers, cached)
  ├── DiskCollector       (/proc/diskstats: per-device throughput, IOPS, latency)
  └── NetworkCollector    (/sys/class/net: per-interface RX/TX, packets, errors)

History (per metric):
  ├── RingBuffer          (short-term: 300 samples for charts)
  └── TimeWindowAggregator (long-term: minute-resolution buckets for 24h stats)

Requirements

• OS: Linux (DGX systems, WSL2, containers)
• GPU: NVIDIA drivers with NVML (libnvidia-ml.so)
• Runtime: No additional dependencies. For GPU monitoring, prefer glibc (-gnu) builds.

Security

dgxtop has undergone a comprehensive security audit addressing:

• UTF-8 boundary safety in all string rendering
• PID recycling protection with pre-kill verification
• Integer overflow protection with saturating arithmetic
• Path traversal prevention in filesystem readers
• Config value sanitization against NaN/Infinity/DoS
• Bounded event channels to prevent memory exhaustion

See commit history for the full audit report and fixes.

License

Apache License 2.0 — see LICENSE for details.

Contributing

Contributions welcome! Please feel free to submit issues and pull requests.