IRSFP
Abstract and scientific scope
IET and VIRR examine the distinction between the number of computational operations and the number of recognised, stable results. The technical discussion distinguishes physical motivation, task-level metrics and empirical benchmark comparisons. Applying these metrics beyond the specific tasks tested requires separate validation.
Performance is not the same as efficiency
A modern processor can perform an enormous number of operations, but some lead to repeated attempts, rejected outputs, corrections or unstable decisions. Conventional OPS/W and TOPS/W metrics describe arithmetic throughput relative to power consumption. By themselves, however, they do not tell us whether energy became a final result that can be distinguished, accepted and retained. This distinction is the starting problem of information-energetic thermodynamics (IET).
A stabilised decision event
IET considers a physically relevant result to be localisable, distinguishable and sufficiently stable, rather than merely transient activity. The full IET formulation assumes a positive entropic cost for a counted event; this condition requires separate physical evidence. This does not reject existing performance measures or Landauer's principle; it complements them by considering the quality and retention of the informational result. On this view, energy-efficient computation depends on whether a system merely generates attempts or produces recognised decisions.
What the public benchmark evaluation found
The practical layer of this programme is called task-level VIRR. In the author's analysis of 27 public MLPerf Tiny v1.3 results with energy-consumption data, the resulting values ranged from approximately 2.73 × 10⁻² to 3.17 × 10⁴ benchmark results per joule. This spread is not, by itself, evidence of a general thermodynamic law: it shows that devices and tasks can differ substantially in their ratios of measurable results to energy. The analysis also included trial variability and bootstrap intervals.
When system rankings may change
A controlled mathematical example shows that ranking devices by performance alone can differ from ranking them by the energy spent on a valid, retained output. Relevant factors include the proportions of accepted and rejected results, valid-output probability, retention requirements, and conversion and control costs. This motivates considering IET and VIRR alongside VSTF methodology, which specifies more precisely when an event can be counted at all.
What the work does not yet establish
Task-level VIRR is a computable benchmark operationalisation, not a measurement of full physical VIRR. Public MLPerf Tiny data do not directly establish local entropy production for each individual decision. Generalising the framework requires independent physical experiments and validation of its assumptions across different system classes.
Papers and public research records
Related public records and identifiers are listed below. A DOI does not, by itself, establish the correctness of findings or guarantee peer review.
- IET/VIRR conference paper — DOI: 10.5121/csit.2026.1601501 ↗
- Supplementary energy–information invariant record — DOI: 10.5281/zenodo.17555326 ↗
Read also
The opening graphic is a conceptual thematic illustration, not an experimental record.