Martin Petrasek, Institute for Resonant Synthesis and Field Physics (IRSFP), Czech Republic

Energy efficiency in information-processing hardware is usually reported through throughput-per-power metrics such as OPS/W or TOPS/W. These metrics count arithmetic activity, but they do not determine whether the physical process has produced a stabilized irreversible informational outcome. We present Information-Energetic Thermodynamics (IET), an operational framework in which the elementary observable is a stabilized irreversible decision event satisfying localization, distinguishability, stability, and positive entropy-production criteria. We distinguish three quantities: full physical VIRR, task-level VIRR, and TOPS/W-derived VIRR estimates. Public MLPerf Tiny v1.3 power submissions demonstrate the computability and cross-submission applicability of the energy-normalized task layer, not all thermodynamic conditions of full VIRR. Across 27 public power results, task-level VIRR spans approximately 2.73e-2 to 3.17e4 benchmark outcomes per joule, with workload-specific summaries, trial variability, and bootstrap intervals. A controlled proof-of-concept analysis then demonstrates mathematically how ranking reversal can occur once valid-output probability, rejection, retention, conversion, and control costs are included. The framework therefore complements TOPS/W by exposing when energy is converted into accepted stable outcomes rather than raw activity.

information thermodynamics, stochastic thermodynamics, irreversible decisions, VIRR, VLSI, MLPerf Tiny, compute-in-memory.