A comprehensive methodology for analyzing and designing energy-efficient neuromorphic systems developed by UNN scientists
Researchers from the Laboratory of Stochastic Multistable Systems at Lobachevsky University have developed a comprehensive methodology for the analysis and design of energy-efficient neuromorphic systems. Alexey Mikhaylov, Director of the Research and Education Centre (REC) "Physics of Solid-State Nanostructures", Ivan Kipelkin, Research Scientist at the Laboratory of Memristor Nanoelectronics, and Grigory Zharkov, Research Assistant at the REC "Physics of Solid-State Nanostructures", worked together with scientists from Russia, Serbia, Italy, the United States, and Spain.
"The key aspect of our work consisted in directly comparing various computing architectures within a unified energy consumption measurement system. Our experiments demonstrated that organic memristors can achieve energy consumption of several femtojoules per event, which is comparable to or even lower than the level observed in biological synapses," stated Ivan Kipelkin, the study lead author from the Laboratory of Memristor Nanoelectronics at the REC "Physics of Solid-State Nanostructures".
For the first time, scientists have proposed a unified metric for comparison: the energy consumption per synaptic event (fJ/event) that enables accurate comparisons between classical neural networks running on GPUs/NPUs, spiking neural networks (Loihi, TrueNorth), memristive devices, as well as biological synapses in rats and humans.
"The key contribution of UNN is the development and study of inorganic memristor crossbars with a 32×8 1T1M architecture. These structures were created within the framework of a state assignment aimed at advancing the electronics industry. At our university, we carried out measurements of currents, calculations of switching energy, and analysis of physical models of neural oscillators based on memristors," noted Alexey Mikhaylov, the project leader andthe Director of the UNN Research and Education Centre"Physics of Solid-State Nanostructures".
As a result of the research, three fundamentally different energy modes were identified: conventional GPUs and NPUs, digital spiking systems, and memristive neuromorphic architectures that closely match the efficiency of biological systems. This is what underpins the necessity of shifting from the traditional von Neumann architecture to in-memory computing and event-driven systems.
"The technologies we have developed can find application in Green AI, autonomous Edge AI devices, robotics, unmanned systems, implants, and sensor networks, where extremely low energy consumption is essential. Moreover, memristor systems hold promise for neuroprosthetics and brain-computer interfaces," emphasised Grigory Zharkov, a research laboratory assistant at the UNN Research and Education Centre"Physics of Solid-State Nanostructures".
The research was performed by Lobachevsky Universityscientists under the Priority 2030 state programme for supporting universities in Russia and the programme of the National Centre for Physics and Mathematics focusing on Artificial Intelligence and Big Data. It was also supported by the Serbian Ministry of Science, Technological Development and Innovation, and the FAIRGROUND project as part of the NextGenerationEU programme. A detailed description of the study by UNN researchers and their colleagues is presented in the article "Critical Analysis of Energy Consumption in Neuro-Computational Systems" published in the IEEE Access journal (Q1).
