Extensive research has been done in the field of resistive switching (or memristive) phenomena in the past decade. By comparison with traditional digital electron-based flash transistors, these emerging devices have multi-state programmability based on ionic redistributions which ensures that once the programming voltage is removed, the desired state is retained without any extra energy consumption. These two characteristics, combined with their simple and scalable design, makes them a promising candidate for implementing artificial synapses in large neuro-inspired hardware. Significant progress has been made at the device level for example, great analog capacity with >100 discernible states per device, very low energy consumption of ~10fJ per programming and very fast switching time of ~85ps. However, there is still no commercial memristive neuromorphic system. Why is that?

This is the question that guides our recent commentary article on memristive crossbars. We got the invitation to write this article in May 2018 as part of a broader series looking at performance vs. manufacturability of emerging technologies. We knew it was going to be hard to write such an article and indeed it took us more than 6 months to have it from idea to print. The challenge was that while a large body of research was done regarding the performance and fabrication at the device level, much less has been achieved at the crossbar level, which requires integration of many devices. The performance and manufacturability challenges faced by memristive crossbars are diverse, but we have decided to focus the three that we believe would make the most significant impact when addressed. These are variability, latency and density. Our paper uses recent results from the literature to highlight the interplay between performance and manufacturability behind each of these challenges. For example, studies have shown that variability decreases in extremely scaled devices, probably due to the confinement of the switching area. However, extreme scaling increases the line resistances which translates in an increased latency.  The trade-offs become even more complicated for large memristor crossbars which typically require compatible integration of the memristor with a non-linear selecting device. The selector in series with each memristor offers increased accessibility. However these selecting devices have their own variability profile and increases the overall variability in the system, thus decreasing the performance. Moreover, they might also reduce the density, thus limiting the realistic crossbar size. We also provide a brief overview of three most common device designs and analyze them from the perspective of manufacturability and their potential advantages / disadvantages in terms of performance.

So far these challenges have prevented industry adoption, but we are optimistic about the future of this technology and its groundbreaking impact on neuromorphic hardware. Therefore, we conclude with a few potential approaches that might help speed up research in the field (Fig. 1). Probably our most significant suggestion is to work as a community to define a data-driven approach towards the device physics, materials and fabrication. When gathering information for our figures, we noticed that it is difficult to corroborate data across different studies and material systems. Appropriate benchmarking and experimental data repositories would strengthen the research across the entire community, from device research all the way to architecture simulations.

Hopefully, our review will support the on-going discussions and help promote community efforts to speed up research progress and technological adoption of memristive systems.

Fig. 1. Roadmap for manufacturing challenges and possible approaches to accelerate progress in the field of memristive neuromorphic hardware.