Holistic Power Modeling for Full System Energy Efficiency

This direction is split into other directions: multi-component energy modeling. Its last version stays readable and citable here.

Version 1 of 178 papers105 conceptsfrozen 2026-10-04

In the quest for sustainable computing, we have historically focused on the "brain" of the machine. Whether it is the Thermal Design Power (TDP) of a CPU or the instantaneous draw of a GPU reported by tools like nvidia-smi, our energy metrics have largely been component-centric. However, a growing body of research suggests that this narrow focus creates a significant "energy blind spot."

Parthasarathy Ranganathan, Suzanne Rivoire, and Christos Kozyrakis argued as early as 2008 in A Comparison of High-Level Full-System Power Models that as systems become more complex, the power consumption of memory and storage subsystems begins to rival or even exceed that of the processor. This research direction advocates for holistic power modeling: a shift from measuring isolated chips to accounting for the entire system-on-chip (SoC), the server chassis, and even the surrounding data center infrastructure.

The Hidden Cost of Auxiliary Components

The intuition behind holistic modeling is simple: a computer is an ecosystem, not a single component. When we run a deep learning model, the GPU does the heavy lifting, but the CPU manages data orchestration, the DRAM stores the weights, and the cooling fans prevent a meltdown.

If we only track the GPU, we miss a massive portion of the energy footprint. Lasse F. Wolff Anthony and the team behind Carbontracker found that neglecting CPUs and DRAM can result in missing up to 50% of the actual energy consumed during training. Similarly, Hao Peng and colleagues, through the Efficiency Pentathlon, demonstrated that during inference, the GPU often accounts for only a minority of total system power.

This "missing energy" isn't just a rounding error; it fundamentally changes which algorithms we perceive as efficient. For instance, Bo Li shows in the Greeness benchmark that the "greenest" model choice depends heavily on usage intensity—a model that looks efficient in a GPU-only vacuum might be an energy hog when the full-cycle costs of retraining and deployment are included.

Fusing Metrics for Accuracy

How do these holistic models work? Instead of relying on a single sensor, they use multi-resource aggregation. By fusing OS utilization metrics (like disk I/O and network activity) with hardware performance counters, researchers can build a more representative picture of energy flow.

  1. Full-Stack Accounting: As described by Keguo Huang in the context of Google's Gemini, this involves aggregating accelerator power with host CPUs, memory, and the "energy tax" of idle machines and cooling. This approach reveals a footprint roughly 2.4 times larger than chip-level consumption alone.
  2. Calibration via Stress Tests: Because software counters like Intel RAPL or Nvidia SMI can be "black boxes" with sampling gaps, researchers like Zeyu Yang and Karel Adámek propose using external hardware (like Power Distribution Units) to calibrate software models. This helps capture the "wall-to-chip" overhead that internal sensors miss.
  3. Marginal vs. Absolute Energy: Akshaya Jagannadharao emphasizes the need to distinguish between the additional power a task draws (marginal) and the total system draw (absolute). In over-provisioned data centers, the absolute energy cost is often much higher than the marginal cost would suggest.

From Microwatts to Megawatts

The need for holistic modeling spans the entire computing spectrum. In the "milliWatt" regime of TinyML, Colby Banbury and Vijay Janapa Reddi argue in TinyMLPerf that measuring kernels in isolation is useless if the firmware and chip peripherals consume the bulk of the power. At the other extreme, Arya Tschand and the MLPerf Power working group are standardizing how we measure megawatt-scale clusters by including interconnects and cooling.

Even in manufacturing, the same logic applies. Chunhui Chung notes that traditional metrics for 3D printing often ignore standby power and gas circulation, leading to incomplete environmental profiles.

Why It Matters: The Regulatory Horizon

This shift toward holistic measurement isn't just academic—it's becoming a regulatory necessity. Alexandra Sasha Luccioni and colleagues have proposed an "AI Energy Star" rating to bring transparency to the field, noting that energy costs can vary by a factor of 16,000 depending on the task and system configuration.

As Kai Ebert and Philipp Hacker point out in their analysis of the EU AI Act, current laws have "green loopholes" because they don't yet mandate these holistic metrics. By moving toward full-system power models, the research community is providing the tools necessary to close those loopholes and ensure that our drive for better performance doesn't come at an invisible environmental cost.

Papers in this version

Showing 60 of 78 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_holistic_power_modeling_for_full_system__98dac3b8,
  title = {Holistic Power Modeling for Full System Energy Efficiency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/holistic-power-modeling-for-full-system-energy-efficiency/rdr_98dac3b8e3fa490ca47abb0e90524ad3},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_holistic_power_modeling_for_full_system__98dac3b8_v1,
  title = {Holistic Power Modeling for Full System Energy Efficiency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/holistic-power-modeling-for-full-system-energy-efficiency/rdr_98dac3b8e3fa490ca47abb0e90524ad3/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}