Skip to main content

Explanation

This section answers the "why" and "how" behind LLenergyMeasure - the questions that sit beneath the practical guides.

Methodology covers why our measurements are scientifically valid: what we measure, how we measure energy, why those choices matter, and how our results relate to other benchmarks. Start here if you want to evaluate or cite the tool in research.

Architecture covers how the system delivers those measurements: the dataflow from configuration to result, the engine-plugin design, the parameter-discovery and curation pipelines, and the CI infrastructure. Start here if you want to understand or extend the internals.

What this section answers​

  • Why is a particular energy measurement approach used? (Methodology)
  • Why does dataset choice affect measurement validity? (Methodology)
  • How does the experiment pipeline convert a config into a result? (Architecture)
  • How are invalid parameter combinations detected before an engine runs? (Architecture)
  • Why does this tool exist and what gap does it fill? (Why)

Sub-sections at a glance​

Methodology​

The scientific grounding for every number LLenergyMeasure produces.

  • Methodology overview - warmup protocol, baseline power subtraction, thermal stabilisation, and reproducibility.
  • What we measure - plain-language guide to energy (joules), throughput (tokens/s), and FLOPs.
  • Energy measurement - NVML telemetry, sampler backends, and known measurement limits.
  • Measurement warnings - what each runtime warning means and how to act on it.
  • Comparison context - how results relate to MLPerf, AI Energy Score, and other benchmarks.
  • Dataset context - why dataset choice shapes energy numbers and guidance for custom datasets.
  • Glossary - definitions for every term used in results, configs, and methodology docs.

Architecture​

The engineering design behind the measurement pipeline.

  • Architecture overview - system components, dataflow, and the harness-plugin contract.
  • Pipeline architecture - the end-to-end path from CLI invocation to written results.
  • CI architecture - how automated engine-coupling validation, schema refresh, and image publishing are wired.
  • Internal pipelines - parameter discovery (runtime validation) and parameter curation (introspection pipeline).

Why​

  • Why LLenergyMeasure - the research gap, what the tool does that is distinctive, and where its design is heading.

Ecosystem​

  • Ecosystem - where LLenergyMeasure sits relative to energy samplers, capability harnesses, and benchmark suites.

Where to start​