The Future is Analog: Energy-Efficient Cognitive Network Functions over Memristor-Based Analog Computations

Current network functions build heavily on fixed programmed rules and lack capacity to support more expressive learning models, e.g. brain-inspired Cognitive computational models using neuromorphic computations. The major reason for this shortcoming is the huge energy consumption and limitation in expressiveness by the underlying TCAM-based digital packet processors. In this research, we show that recent emerging technologies from the analog domain have a high potential in supporting network functions with energy efficiency and more expressiveness, so called cognitive functions. We propose an analog packet processing architecture building on a novel technology named Memristors. We develop a novel analog match-action memory called Probabilistic Content-Addressable Memory (pCAM) for supporting deterministic and probabilistic match functions. We develop the programming abstractions and show the support of pCAM for an active queue management-based analog network function. The analysis over an experimental dataset of a memristor chip showed only 0.01 fJ/bit/cell of energy consumption for corresponding analog computations which is 50 times less than digital computations.