GPU & CPU Monitor

Your machine's vital signs, one glance away.

Stop alt-tabbing to a terminal to run nvidia-smi and htop.
GPU & CPU Monitor lives in your VS Code sidebar and shows every GPU, every core, and the exact processes (and models) behind the load, live, while you work.

Why you'll like it

You are training a model, serving an LLM, or just wondering why the fans are screaming. The answer is on your machine, and getting to it should cost you nothing. This extension turns the two commands you keep retyping into a calm, glanceable panel that sits right next to your code.

• One look, full picture. GPU load, VRAM, memory bandwidth, power, temperature, per-core CPU load and RAM, all in one place.
• It knows what's running. Processes on your GPUs show the model name if they belong to an SGLang or vLLM server, not a cryptic python3.
• It stays out of your way. Light on resources, quiet on screen, and it adapts to any sidebar width.

The GPU tab

A card for every GPU, with utilization, VRAM, memory bandwidth and power as clean bars that shift from green to yellow to red as load climbs, plus used and total VRAM and the live temperature.

A compute process table that tells you what is actually going on:

Tip: click any process row to copy its PID to your clipboard. Handy for the kill you were about to type.

How model names are found

SGLang and vLLM run their work in child processes whose names say very little. The extension walks up the process tree to the server that launched them and reads its arguments, preferring --served-model-name, then --model-path / --model, then vllm serve <model>. The result is cached, so it is effectively free.

The CPU tab

A card for every physical processor. Dual-socket server? You get two cards, each with its own CPU utilization, its own RAM (read per NUMA node, so the numbers are honest), and its package temperature.

Click a card to open the cores. A smooth, eased animation reveals every logical core as a round button showing its real core number, colored by load:

Colors cross-fade as load changes (and the circles can become boxes, or any colors you like, see Make it yours), and hyperthread siblings (for example 0-23 then 48-71) are drawn as separate runs with a faint divider, so the numbering always makes sense at a glance.

Top processes, with the user who owns each one and the same model-name detection as the GPU tab. Utilization is per core, like top, so a busy multi-threaded process can exceed 100%. Click a row to copy its PID.

Make it yours

The look of the core buttons lives in a plain JSON file you control. Point the extension at it:

// settings.json
"nvidiagpumonitor.appearanceFile": "~/gpu-monitor.json"

{
  "core": { "shape": "circle", "size": 28, "radius": 6, "gap": 5 },
  "levels": [
    { "max": 0,   "color": "#6b6b6b", "text": "#dddddd", "opacity": 0.6 },
    { "max": 24,  "color": "#3fd68f", "text": "#ffffff" },
    { "max": 49,  "color": "#e8d44d", "text": "#000000" },
    { "max": 74,  "color": "#d18616", "text": "#ffffff" },
    { "max": 99,  "color": "#f2555a", "text": "#ffffff" },
    { "max": 100, "color": "#8c2a2e", "text": "#ffffff" }
  ]
}

Edits apply live, with no reload. Anything missing or invalid quietly falls back to the defaults, so a typo can never break the panel. A ready-to-copy file ships as appearance.example.json. Relative paths are resolved from your workspace folder, and ~/ works too.

Beautiful at every size

Narrow the sidebar and the panel does not crumble. It simplifies. Each stage collapses with one smooth easing curve instead of snapping:

1. Full: bars, names and every stat.
2. Compact: totals fold away and you keep what is used, plus the temperature.
3. Minimal: bars give way to live percentages, names step aside, and the card is centered around its GPU / CPU ID.

Columns that carry no information are removed for you. A single-GPU or single-CPU machine will not see a pointless index column.

Gentle on your machine

A monitor should never become the thing you are monitoring.

• Only the tab you are looking at is sampled.
• Everything runs asynchronously, so the editor never waits on nvidia-smi.
• A refresh is skipped if the previous one is still running, so slow reads cannot pile up.
• Hidden panel? Polling pauses until you open it again.

Requirements

Developed and tested on Linux. Other platforms are untested.

Install

From a VSIX

code --install-extension gpu-monitoring-service-0.2.0.vsix

From source

npm install
npm run compile
npx vsce package          # builds the .vsix
code --install-extension gpu-monitoring-service-0.2.0.vsix

Then open the GPU & CPU Monitor icon in the activity bar.

Working over Remote - SSH? Install it on the remote host and the panel shows that machine's GPUs and CPUs, which is usually exactly what you want.

Settings

Command: GPU & CPU Monitor: Refresh Now forces an immediate update.

Privacy

Everything happens locally. The extension makes no network requests and collects no telemetry. To recognize SGLang and vLLM models it reads process command lines, extracts only the model name, and keeps nothing.

Troubleshooting

Something not right? Open View → Output and pick GPU & CPU Monitor for a log of every failure (raw error text included). Problems also show up in the panel itself, in plain language.

"nvidia-smi error" banner. The executable was not found or the driver is not responding. Run nvidia-smi in a terminal, or set the full path in nvidiagpumonitor.nvidiaSmiPath.

Process utilization always shows -. Your driver may not support nvidia-smi pmon. Everything else keeps working.

A process shows python3 instead of a model. Only SGLang and vLLM launches are recognized, and only when the server's command line includes the model argument.

No CPU temperature. Your machine has no Intel coretemp sensor exposed under /sys/class/hwmon.

_{The screenshots above use sample data so no real machine details are shown.
Built for people who like to see what their hardware is doing.}