Ultra-low power logic in memory with commercial grade memristors and FPGA-based smart-IMPLY architecture

Reducing power consumption in nowadays computer technologies represents an increasingly difficult challenge. Conventional computing architectures suffer from the so-called von Neumann bottleneck (VNB), which consists in the continuous need to exchange data and instructions between the memory and the processing unit, leading to significant and apparently unavoidable power consumption. Even the hardware typically employed to run Artificial Intelligence (AI) algorithms, such as Deep Neural Networks (DNN), suffers from this limitation. A change of paradigm is so needed to comply with the ever-increasing demand for ultra-low power, autonomous, and intelligent systems. From this perspective, emerging memristive non-volatile memories are considered a good candidate to lead this technological transition toward the next-generation hardware platforms, enabling the possibility to store and process information in the same place, therefore bypassing the VNB. To evaluate the state of current public-available devices, in this work commercial-grade packaged Self Directed Channel memristors are thoroughly studied to evaluate their performance in the framework of in-memory computing. Specifically, the operating conditions allowing both analog update of the synaptic weight and stable binary switching are identified, along with the associated issues. To this purpose, a dedicated yet prototypical system based on an FPGA control platform is designed and realized. Then, it is exploited to fully characterize the performance in terms of power consumption of an innovative Smart IMPLY (SIMPLY) Logic-in-Memory (LiM) computing framework that allows reliable in-memory computation of classical Boolean operations. The projection of these results to the nanoseconds regime leads to an estimation of the real potential of this computing paradigm. Although not investigated in this work, the presented platform can also be exploited to test memristor-based SNN and Binarized DNNs (i.e., BNN), that can be combined with LiM to provide the heterogeneous flexible architecture envisioned as the long-term goal for ubiquitous and pervasive AI.