pi-gpu-cost

Track GPU power draw and estimated electricity cost for each pi session.

Its purpose is simple: pi runs on the same machine as the GPU. The extension samples the local nvidia-smi on the machine where the pi process runs and keeps its log there too — there is no remote mode, so a pi running somewhere else (e.g. on a laptop pointed at a remote GPU box) cannot see this GPU.

⚠️ Requirements — read this first

This package samples power with nvidia-smi, so it requires all of:

1. pi on the same machine as the GPU. The extension spawns nvidia-smi locally on the machine where the pi process runs. A pi running elsewhere will find no GPU and tracking is disabled (with a warning, no crash).
2. An NVIDIA GPU (AMD, Intel, and Apple Silicon are not supported — there is no power-draw backend for them).
3. nvidia-smi on your PATH. It ships with the NVIDIA driver — install or update your driver if you don't have it.

Verify before installing:

nvidia-smi --query-gpu=power.draw --format=csv,noheader,nounits
# -> a number like "242.15" means you're good
# -> "command not found" / an error means the driver (and nvidia-smi) is missing

If nvidia-smi is unavailable, the extension loads but logs gpu-cost: nvidia-smi not available — GPU cost tracking disabled and shows no footer. It never crashes the session.

What it does

Samples nvidia-smi every few seconds while a pi session is running, integrates power over time into energy (Wh), and turns that into money using your electricity rate.

• Live footer status: ⚡ 269W · 97% · 61°C · 14.2GB · 0.12 kWh · €0.03
• /gpucost — current session (avg/peak watts, utilization, energy, cost), live GPU status (name, VRAM used/total, temperature), the processes using the GPU, today's total across closed sessions, and a monthly cost overview (newest months first)
• On quit: summary notification + one JSON line per session in log.jsonl and one row in the semicolon-separated log.csv (avg/peak watts, peak temperature, energy, cost)

No data ever leaves your machine.

Install

pi install npm:pi-gpu-cost          # npm (once published)
pi install git:github.com/heinbit/pi-gpu-cost   # git
pi install ./path/to/pi-gpu-cost    # local path

Then start (or /reload in) a pi session.

Set your electricity rate (do this once)

The cost column is only as good as your rate. The built-in default is $0.30/kWh — set your real price so the numbers mean something.

1. Create your config file in the data directory:

mkdir -p ~/.pi/gpu-cost
cp config.example.json ~/.pi/gpu-cost/config.json   # from the package folder, or just create it

2. Edit ~/.pi/gpu-cost/config.json and set your rate:

{
  "ratePerKwh": 0.282,
  "currency": "€",
  "intervalMs": 5000,
  "idleWatts": 0
}

ratePerKwh is the price of one kilowatt-hour in your currency — the same number your electricity bill uses. Examples:

You bill in…	ratePerKwh	currency
€0.282 / kWh (DE)	0.282	"€" or "EUR"
$0.16 / kWh (US)	0.16	"$" or "USD"
£0.30 / kWh (UK)	0.30	"GBP"
28.2 ct / kWh	0.0282	"€"

Both work: ISO 4217 codes (EUR, USD, GBP, JPY, PLN, …) are converted to their symbol; any other string ("€", "kr", "fr") is shown as-is.

That's it — costs update from the next session on (env overrides apply live).

All config keys

key	default	meaning
ratePerKwh	0.3	electricity price per kWh (your main setting)
currency	"$"	ISO 4217 code ("EUR", "JPY", …) or any symbol ("€", "kr")
intervalMs	5000	nvidia-smi sampling interval (min 50)
idleWatts	0	idle baseline subtracted from draw (0 = no offset)

Precedence & overrides

Config is resolved in this order (first hit wins):

1. Environment variables: GPU_COST_RATE, GPU_COST_CURRENCY, GPU_COST_INTERVAL_MS, GPU_COST_IDLE_WATTS
2. ~/.pi/gpu-cost/config.json (your personal config, stable across updates)
3. config.json next to the extension (if shipped with the package)
4. Built-in defaults ($0.30/kWh)

GPU_COST_DIR moves the whole data directory (default ~/.pi/gpu-cost).

Notes

• Energy = ∫(power.draw − idleWatts) dt over the session. Failed nvidia-smi ticks extrapolate at the last known wattage for up to 30 s, so brief hiccups don't undercount.
• Multi-GPU: power and VRAM are summed across GPUs, utilization is the max.
• idleWatts: set it to your idle draw (e.g. 25) if you only want the cost attributable to active work, not the baseline.

Data

log.jsonl in the data directory holds one JSON line per closed session:

{"sessionId":"…","startedAt":"…","endedAt":"…","durationMin":39.7,"samples":476,"failedSamples":0,"avgWatts":215.1,"peakWatts":271.9,"avgUtil":75,"energyWh":142.531,"cost":0.04019,"ratePerKwh":0.282,"reason":"quit"}

A legacy log.jsonl from older versions (stored next to the extension) is migrated automatically on first use.

log.csv in the same directory is the long-term log: one semicolon-separated row per closed session (same fields as above), CRLF line endings, written on every session close — open it directly in Excel or any spreadsheet tool:

sessionId;startedAt;endedAt;durationMin;samples;failedSamples;avgWatts;peakWatts;peakTempC;avgUtil;energyWh;cost;ratePerKwh;reason
abc123;2026-08-16T06:00:00.000Z;2026-08-16T06:39:42.000Z;39.7;476;0;215.1;271.9;85;75;142.5;0.04019;0.282;quit

If log.csv does not exist yet, it is backfilled from log.jsonl on startup, so the CSV covers your full history from day one. (Decimals use . as the separator; fields containing ; or quotes are quoted.)

/gpucost also prints a monthly overview of closed sessions (newest months first, up to 6 months; each session is counted at the rate that was active when it ran):

monthly (closed sessions, newest first):
  2026-08: 1.42 kWh = €0.4004 (12)
  2026-07: 0.85 kWh = €0.2397 (7)

Development

npm install
npm test         # unit + sampler lifecycle + real nvidia-smi smoke test (skipped if no GPU)
npm run typecheck

Tests use Node's built-in test runner with native TypeScript type stripping (Node ≥ 22.18) — no test framework dependency.