Thermodynamic Constraints as Organizing Principles of Neural Efficiency and Cognitive Architecture.

Biological neural systems sustain adaptive cognition under persistent and severe metabolic constraints. This paper develops a multi-scale thermodynamic framework in which those constraints function as organizing principles of neural efficiency and cognitive architecture, addressing four distinct shortcomings of prior accounts. First, I replace the widely cited "2 bits per synapse" convention with a probabilistic model grounded in the Bernoulli statistics of downstream neuron firing, yielding a substantially more empirically defensible efficiency ratio of approximately 3.4 × 106 times the Landauer limit. Second, I close the quantitative gap between per-synapse ATP chemistry and the brain's globally measured 20 W power budget through an explicit multi-scale bridge. Third, sparse coding, predictive processing, and cross-frequency coupling are derived as Lagrangian solutions to a single metabolic optimisation functional rather than merely described as consistent with energetic principles. Fourth, neuromorphic efficiency comparisons are updated and standardised to an energy-per-bit metric for Intel Loihi 2, BrainScaleS-2, and SpiNNaker2, and are extended to two compute-in-memory architectures - the charge-recycling array processor of Karakiewicz et al. (2012) and the RRAM-based NeuRRAM chip of Wan et al. (2022) - both of which approach or exceed biological synaptic efficiency. Each coding strategy generates testable, quantitatively specified predictions that admit principled rejection. The framework positions metabolic pressure not as an engineering detail extrinsic to neuroscience but as a constitutive selective force in the evolutionary shaping of neural computation.