Linux GPU Monitor

This tool provides Linux-focused GPU utilization monitoring with per-GPU granularity. It targets system administrators, DevOps engineers, ML engineers, and performance analysts who need consistent visibility into GPU workloads, memory pressure, temperatures, and power draw. By collecting metrics from standard interfaces, it avoids vendor lock-in while enabling cross-vendor comparisons and capacity planning.
Conceptually, the tool runs a configurable sampler that queries GPU drivers at fixed intervals and emits timestamped records. It supports NVIDIA via nvidia-smi and AMD via ROCm, with optional fallbacks to generic /proc data when available. Outputs are designed for ingestion by data pipelines, dashboards, or alerting rules.
Unique value lies in its portability, strict per-GPU reporting, and simple export formats. It provides consistent units (percent, MB/GB, Celsius, watts), handles data gaps gracefully, and can be integrated with existing monitoring stacks. Use cases include training queue management, batch job optimization, and proactive thermals/anomaly detection in shared workstations or data centers.
Core features include multi-GPU support, configurable sampling, per-GPU utilization/memory/temperature/power, ISO 8601 timestamps, JSON/CSV outputs, and optional aggregations. Typical workflows involve baseline characterization, threshold tuning, and exporting to dashboards for SLA and capacity planning.