Measure the emissions of compute you run yourself by sampling hardware power and applying local grid carbon intensity

The proposal

Wrap training, inference or batch jobs on hardware you control in a tracker that samples power draw over the run, converts it to energy, and multiplies by the grid's CO2 intensity for that place and time to produce a logged emissions estimate.

Why it works

Energy times carbon intensity is the direct physical basis of operational emissions. Because you own the hardware, power can be sampled rather than guessed, using interfaces such as Intel RAPL for CPU and DRAM and vendor tools like NVIDIA NVML for GPU, so the main uncertainty sits in the grid-intensity factor rather than the energy figure.

Where it fits

Self-hosted or cloud compute where hardware counters are readable. Strongest for ML training and inference, but applies to any long-running job. In orchestrated environments the same mechanism can be pushed down to per-container and per-pod granularity.

Trade-offs and limitations

It is an estimate, not metered ground truth. Grid-intensity data quality varies by region and time resolution. Implementations typically capture the main compute components (CPU, GPU, RAM) and do not, by default, model disk, network, cooling or data-centre PUE, nor the embodied emissions of manufacturing the hardware. Instrumentation adds small overhead, and where kernel-level sampling is used it can miss very short-lived processes.