전문가처럼 nvidia-smi 읽기

Think of nvidia-smi as the dashboard for your GPU. One command shows the model, memory, temperature, and what is running right now. 📊

The very top line shows your driver version and the maximum CUDA it supports. Check here first whenever a version error appears.

The first table row lists each GPU's name and an index number. With several cards, those indices are how you pick which device to use.

The Memory-Usage column shows used versus total VRAM. If it is nearly full, your next allocation may fail with an out-of-memory error.

The GPU-Util percentage tells how busy the cores are. A stuck-at-zero number while your kernel runs usually means the work never reached the device.

The Temp and power columns reveal thermal health. Sustained high temperatures can cause the GPU to throttle and quietly slow your kernels.

The bottom Processes table lists every program currently using the GPU, with its PID and memory footprint. Great for spotting a forgotten job.

Add a flag to refresh the dashboard every second, so you can watch memory and utilization change while your program runs.

For scripts, query mode prints just the fields you ask for as clean, parseable text instead of the full visual table.

On data-center cards nvidia-smi also reports ECC error counts and persistence mode. For laptops and desktops these fields are usually off or absent.

With many cards, the -i flag focuses the report on a single GPU by index. Handy when you only care about the device your job is using.

Let us read the dashboard together.

You can now read nvidia-smi like a pro: driver line up top, memory and utilization in the table, and live processes at the bottom. A glance tells the whole story. ✅