Power usage: see how much electricity your AI actually uses
Running your own inference means running your own hardware, and hardware uses electricity. Until now the console could tell you how many requests a worker served and how fast, but not what it cost the machine to serve them. From today every worker reports how much power its machine is drawing. You can watch it live on the worker's page and, on Pro, look back over the last hour, day, week or month.
Some teams want to understand energy usage for their reporting. Others want to know whether a GPU is running flat out, or how much power and cooling a second machine would need. Either way, we've got you covered.
What you'll see
Open Workers and each worker shows its live draw beside its status, such as 512 W. On the worker's own page, a Power tile sits with the other vitals. It shows the whole machine's draw right now, refreshed every few seconds, and on NVIDIA hardware a bar for how hard the GPUs are working against their power limits.
Further down the page, the Power usage card charts that draw over time. The chart is stacked by part, so you can see how much of the total is each GPU, how much is the CPU and how much is everything else in the box. It shows kWh usage, average and peak draw, and how much of the period has readings.
A worker reports its power usage every few seconds. Different time periods allow you to see peaks, troughs and averages.
Measured or estimated
Very few machines can report their full draw from a single sensor, so every figure says how it was arrived at. A Mac with Apple silicon reports the whole machine from its own power sensor, and Pendra labels it Measured. On a PC or server, each NVIDIA GPU is read from the driver and, on Linux, the CPU from its own energy counter. The motherboard, memory, fans, drives and power-supply losses usually have no sensor at all, so Pendra estimates their share, adds it on and labels the total Estimated. Hover the label to see exactly which parts were measured.
On a virtual machine, such as a cloud GPU instance, the worker can see its GPUs but not the host around them, so the estimate adds a fixed allowance for each GPU. If a machine has no sensor the worker can read, it reports nothing and the console says so. We would rather show a gap than a number we made up.
What leaves your machine
The worker sends one-minute averages for each part it reads (average, lowest and highest draw, plus GPU load and temperature) and a live reading every few seconds for the tile. Power reporting collects nothing about individual requests, prompts or responses.
It is on by default. If you would rather a machine didn't report it, set
PENDRA_DISABLE_POWER_TELEMETRY=1 where the worker runs and
restart it. The console then shows that power reporting is off, and
anything already recorded stays in the history.
Which plan gets what
The live draw, on the worker's page and in the Workers list, is on every
plan, Free included. The history is part of Pro and
Enterprise. That covers the chart from one hour to 30
days, the energy, average and peak for each period, and the same data
through the API at GET /api/v1/workers/{worker_id}/power
for your own dashboards. Workers report their readings whatever your plan,
so if you upgrade later, the chart already covers the weeks before.
Enterprise customers who need energy figures in a particular shape, for a sustainability report or an internal cost model, can ask us for custom energy reporting built on the same data.
Power usage works on any worker from version 3.112 onwards. Just open a worker in the console to see it. Find more technical information in the docs.
If you want to talk to us about energy reporting, get in touch.