Measurement of Generative AI Workload Power Profiles for Whole-Facility Data Center Infrastructure Planning

2026-04-08 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure · Depth: Advanced, medium

Summary

A new methodology measures the power consumption of generative AI workloads at 0.1-second resolution, linking these high-resolution measurements to whole-facility energy demand for data center infrastructure planning. Researchers used NLR's high-performance computing data center, equipped with NVIDIA H100 GPUs, to profile AI training, fine-tuning, and inference jobs. Workloads were characterized using MLCommons benchmarks for training and fine-tuning, and vLLM benchmarks for inference, ensuring reproducibility. The collected power profiles are then scaled to the whole-facility level using a bottom-up, event-driven energy model. This approach captures realistic temporal fluctuations driven by AI workloads and user behavior, providing critical data for planning grid connections, on-site energy generation, and distributed microgrids. The dataset of power consumption profiles is publicly available.

Key takeaway

For data center infrastructure planners and CTOs evaluating future energy needs, understanding the precise power demands of generative AI workloads is crucial. Your teams should integrate high-resolution AI power profiles into facility planning to accurately forecast energy consumption and optimize infrastructure investments for grid connections, on-site generation, and microgrids. This data-driven approach mitigates risks associated with under- or over-provisioning.

Key insights

High-resolution AI workload power profiles enable accurate whole-facility data center energy planning.

Principles

• Standardized benchmarks ensure reproducible workload profiling.
• Bottom-up modeling scales workload data to facility level.

Method

Measure AI workload power at 0.1-second resolution using NVIDIA H100 GPUs, characterize with MLCommons/vLLM benchmarks, then scale profiles to whole-facility energy demand via an event-driven model.

In practice

• Utilize public dataset for AI power consumption.
• Inform grid connection planning with temporal power profiles.

Topics

• Generative AI Workloads
• Data Center Energy Planning
• Power Consumption Profiling
• NVIDIA H100 GPUs
• MLCommons Benchmarks

Code references

• erfan-nourbakhsh/GenAI-EdSent

Best for: CTO, VP of Engineering/Data, Director of AI/ML, MLOps Engineer, AI Architect, AI Hardware Engineer

Related on AIssential

• A faster way to estimate AI power consumption — EnergAIzer rapidly and accurately estimates AI workload power consumption on GPUs, improving data center efficiency.
• Running AI Workloads on Rack-Scale Supercomputers: From Hardware to Topology-Aware Scheduling — NVIDIA Mission Control and related tools bridge hardware topology with scheduler abstractions for optimal AI/HPC workload placement.
• Gartner: Data Centres Face 26% Energy Spike From AI Servers — Rapid AI server growth is creating an unprecedented energy deficit, making power availability a critical constraint for data center expansion.
• As AI data centers hit power limits, Peak XV backs Indian startup C2i to fix the bottleneck — Efficient power delivery is critical for scaling AI data centers and reducing operational costs.
• AI and Power Consumption Explained — AI's rapid growth drives substantial increases in global electricity and water consumption, posing significant environmental challenges.
• AMD Soars on Blockbuster AI-Fueled Forecast — AI demand is driving significant growth and strategic shifts across semiconductor, cloud, and consumer tech sectors.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.