Why it matters

nvidia-smi is the first tool anyone reaches for. Understanding it is basic GPU literacy.

Advertisement

The architecture

Default view: table of GPUs with util, mem, temp, power, processes.

Query mode: --query-gpu, --query-compute-apps for scripted use.

nvidia-smi viewsDefault tablequick status--query modesscripted output-l Nloop every N secGPU utilization is misleading: 100% means at least one SM busy; use DCGM for detail
Common uses.
Advertisement

How it works end to end

Common flags: -l for loop, --query for structured, --format for output format.

Persistence mode: --persistence-mode 1 keeps driver loaded (faster launches).

Compute mode: exclusive vs default (multi-process).

ECC control: view + reset ECC error counters.