Critical Analysis of Energy Consumption in Neuro-Computational Systems

This paper presents a unified benchmark of energy consumption per synaptic event across five classes of neuro-computational substrates: graphics processing units, neural processing units, field-programmable gate arrays, digital spiking processors, and in-memory/memristive devices, with biological synapses used as reference bands. The main contribution is methodological as well as empirical. Methodologically, we introduce a common event-level metric that makes heterogeneous systems directly comparable despite major differences in architecture, coding scheme, and learning dynamics. Empirically, we combine measurements obtained by the authors with carefully normalized literature data to map the present energy landscape of artificial and biological neural computation. The results reveal three robust regimes. Dense von Neumann ANN implementations on GPUs, NPUs, and FPGA operate mainly in the 105 to 108 fJ range per synaptic event. Digital spiking processors reduce this requirement to about 103 to 104 fJ per event. Memristive and transistor-based artificial synapses span a much broader interval, from array-level values near 106 fJ down to a few femtojoules in the most efficient device-level realizations. In particular, the best organic PANI memristive samples approach ∼ 2 fJ per synaptic event, entering the biological vicinity and, in some cases, surpassing the rat reference range. Taken together, the benchmarked landscape spans up to ten orders of magnitude across the full set of systems considered in this study. This result clarifies where current AI hardware stands relative to biological efficiency, identifies event-driven and in-memory computation as the most promising routes toward sustainable AI, and provides a quantitatively grounded reference framework for future neuromorphic benchmarking.