Dgxtop

dgxtop v0.2.1 — Comprehensive GPU Detail & Bandwidth Monitoring

New: Full-Page GPU Detail View

The GPU Detail tab now shows a single selected GPU with all available NVML metrics in an organized layout:

Utilization & Memory

• GPU %, VRAM used/total, Memory Bandwidth (actual/theoretical GB/s), BAR1 usage

Thermal & Power

• Temperature (with slowdown/shutdown thresholds), fan speed
• Power draw/limit with total energy consumption (kWh)
• Throttle reasons decoded: Idle, SwPowerCap, HwSlowdown, SwThermal, HwThermal, HwPowerBrake, etc.

Clocks

• All 4 domains: Graphics, SM, Memory, Video (current MHz)

PCIe & Connectivity

• PCIe Gen/Width (current vs max)
• TX/RX throughput
• NVLink topology (active links → remote GPU mapping)
• Encoder/Decoder utilization

Health & Info

• ECC errors (corrected + uncorrected)
• Retired pages (SBE/DBE)
• Performance state (P0–P15), Compute mode, Persistence mode
• UUID

Per-GPU Processes

• Processes filtered to the selected GPU (PID, GPU%, VRAM, command)

Charts

• 4 time-series charts: Utilization, Memory, Temperature, Power
• 1/6/12/24h historical stats

VRAM Bandwidth Monitoring

MemBW   55% util ~850/1555 GB/s

Calculates theoretical peak bandwidth from memory_clock × bus_width × 2(DDR) / 8 and actual bandwidth from utilization %. Useful for identifying if memory bandwidth is the bottleneck during training — especially relevant for DGX Spark (GH200) vs consumer GPUs.

README Updated

• Feature tables updated with all new GPU metrics
• New CLI option --net-max documented
• Keyboard shortcuts updated with n/N (network) and d/D (disk)
• Architecture diagram updated with new collectors and history system

Install

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

Full Changelog: https://github.com/DennySORA/dgxtop/compare/v0.2.0...v0.2.1

Full Changelog: https://github.com/DennySORA/dgxtop/compare/v0.2.0...v0.2.1