Generative AI Workload Power Profiles at 0.1-Second Resolution

High-resolution power consumption traces for generative AI workloads executed on the Kestrel HPC platform. The dataset includes profiles for inference and training jobs across LLM and image generation tasks, with varying datasets and node counts. Published by the Department of Energy in 2026, it aims to address the lack of open empirical energy data for AI systems.

Use Cases

• Data center infrastructure planning based on workload-specific power profiles
• Energy modeling and validation of simulation tools using the bottom-up, event-driven model examples
• Demand response and grid impact studies based on time-resolved electrical power measurements
• Sustainability analysis for large-scale AI deployment using empirical energy data

Strengths

• Power measurements are captured at high resolution (5/10 Hz, or every 0.2/0.1 seconds)
• Profiles cover diverse configurations including job type, workload, datasets, and node count
• Each profile is accompanied by metadata describing workload parameters for reproducibility
• Data is provided in both raw/instantaneous and aggregated formats

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment
• Data format is BIN, which may require specialized tools for access

Provenance

Source

Department of Energy, measurements from NLR's Kestrel HPC platform

Collection Method

Time-resolved electrical power measurements and bottom-up, event-driven modeling

Time Range

null

Freshness

Last updated 2026-04-11 16:02:52.105228

Geography

null