Training Carbon Footprint Calculator
Inputs
| GPU-hours | 1,000,000 |
|---|---|
| Power per GPU | 700 W |
| Power usage effectiveness | 1.1 |
| Carbon intensity | 400 |
Training Carbon Footprint Calculator
Estimate the electricity and carbon emissions of an LLM training run from total GPU-hours, per-GPU power draw, the data center’s power usage effectiveness, and the grid’s carbon intensity.
Inputs
Compute
Grid
Results
Enter a value to see results.
Details
Training Carbon Footprint
A large training run draws steady power for weeks, and the carbon it emits is that electricity multiplied by how dirty the grid is. This calculator estimates the operational footprint of a run from the total GPU-hours, the average power each accelerator draws, the data center's power usage effectiveness, and the carbon intensity of the electricity supply.
The quantities counted
The footprint here is operational: the emissions from the electricity consumed while training, not the embodied carbon of building the chips or the facility. The energy a run uses is the accelerator-hours times the average power per accelerator, which gives the energy the GPUs themselves draw. Real facilities also spend energy on cooling, power conversion, and distribution, so that figure is scaled up by the power usage effectiveness. Multiplying the resulting electricity by the grid's carbon intensity converts kilowatt-hours into kilograms of CO2.
The formula
With GPU-hours, an average power per accelerator in watts, a power usage effectiveness of PUE, and a grid carbon intensity in grams of carbon dioxide per kilowatt-hour, the energy and emissions are
EM=1000H⋅p⋅PUE=1000E⋅IDividing power in watts by 1000 yields kilowatts, so comes out in kilowatt-hours; dividing the grams by 1000 yields kilograms, and a further division by 1000 gives metric tonnes. Every input enters linearly, so halving any one of them halves the footprint.
Worked example
Take a run of one million GPU-hours, each accelerator drawing 700 watts on average, in a facility with a power usage effectiveness of 1.1, on a grid at 400 grams of CO2 per kilowatt-hour:
EM=10001,000,000×700×1.1=770,000 kWh=1000770,000×400=308,000 kg=308 tonnesThe run consumes 770 megawatt-hours and emits 308 tonnes of CO2. Moving the same run to a grid at 100 grams per kilowatt-hour would cut the emissions to 77 tonnes without changing the energy at all, which is why location and scheduling often dominate the footprint.
Limits
The estimate covers only operational electricity emissions. It excludes the embodied carbon of manufacturing the hardware and constructing the facility, which can be significant for short runs on new equipment. Power draw varies with workload and is rarely at the rated maximum, power usage effectiveness changes with climate and load, and grid carbon intensity shifts hour to hour as the generation mix changes. Use measured power and real-time grid figures where available. The ongoing energy cost of serving the trained model is covered in the Inference Energy Cost Calculator, and the rental side of the same compute in the GPU Cloud Cost Calculator.
Frequently Asked Questions (FAQ)
How are the emissions estimated?
The run’s electricity use is computed first: the accelerator-hours times the average power per accelerator give the GPU energy, and multiplying by the power usage effectiveness adds the cooling and distribution overhead of the facility.
That total energy in kilowatt-hours is then multiplied by the grid’s carbon intensity in grams of CO2 per kilowatt-hour. The result is the operational carbon footprint — the emissions from the electricity consumed while training, expressed in kilograms and tonnes.
What is power usage effectiveness?
Power usage effectiveness, or PUE, is the ratio of the total energy a data center draws to the energy that actually reaches the computing equipment. The excess goes to cooling, power conversion, and distribution.
A PUE of 1.5 means every watt of compute carries half a watt of overhead, while an efficient hyperscale facility reaches about 1.1. Because the figure multiplies the whole energy estimate, choosing a value that matches the actual facility matters as much as the GPU power itself.
Why does grid carbon intensity matter so much?
Carbon intensity is the emissions per unit of electricity, and it varies enormously by location and time. The same run on a coal-heavy grid above 700 grams per kilowatt-hour can emit several times more than on a grid rich in hydro, wind, or nuclear below 100.
Because emissions scale directly with this figure, where and when a run is scheduled often changes its footprint more than any efficiency tweak. Annual averages give a planning estimate; real-time grid data gives a sharper one.
Disclaimer
This estimate covers only operational emissions from electricity use and excludes the embodied carbon of manufacturing the hardware and building the facility. Power draw, power usage effectiveness, and grid carbon intensity vary over time and location; use measured values where available and treat the result as an approximation.