In state-of-the-art von-Neumann computing architectures, the central processing unit (CPU) and the memory are physically separated and communicate only via a finite-bandwidth data bus. In the Internet-of-Things and Big Data processing era, however, the data transfer becomes the limiting factor in computing performance. This limitation, also called the von-Neumann bottleneck, poses a serious constraint for novel popular applications requiring a great deal of data transfer. Thus, the research community is exploring new computing paradigms, in which memory and processing unit are integrated one onto the other.

One promising approach is based on the Cellular Nonlinear Network (CNN) computing paradigm. A single CNN cell consists of a capacitance, a resistance, and a number of voltage-controlled current sources. The cells are organized in a 2D or 3D grid, in which only neighboring cells are physically coupled. The network performs different computation tasks on the basis of the coupling and self-coupling arrangements. Back in 1993, T. Roska proposed to add local storage units to each CNN cell to realize a Universal Machine (UM), the first ever demonstration of a non-von Neumann computer.

CNN cells have been implemented in hardware in combination with photodetectors for the development of very efficient visual microprocessors. The integration of memory units into each cell, however, consumes a large space, and, finally, leads to the poor spatial resolution CNN-based vision sensors suffer from. In order to overcome this issue, very simple non-volatile memory elements with high scalability are required. Memristive devices, especially based on redox mechanisms, show such properties and, thus, they could be the ideal technology candidate to resolve the area-based performance limitation of state-of-the-art CNN-UMs.So far, CNN-UMs have been analyzed only on a theoretical and simulation domain, while a verification in hardware is still missing.

In the proposed Mem2CNN project, the feasibility of Memristor CNN-UM (MCNN-UM) is explored. To this end, the group at TU Dresden teams up with the group at FZ Jülich, combining expertise on memristor and CNN theory (TU Dresden), and on memristor fabrication and physical modeling of memristive devices (FZJ). In the project, the physical compact models of FZJ will be reformulated to enable a more fundamental theoretical framework for memristive circuit modeling and investigation. Moreover, single M-CNN cells will be fabricated and characterized for the first time. For this purpose, a small hardware demonstrator, consisting of a matrix of 6×6 MCNN cells, will be showcased to demonstrate the execution of various tasks, such as the extraction of corners or edges from an input image. Based on the theoretical framework for MCNNs, larger networks will be designed and simulated to demonstrate the potential for computation universality, versatility, and efficiency of the MCNN-UM memcomputing approach.