Training and recognition with neural networks generally require high throughput, high energy efficiency, and scalable circuits to enable artificial intelligence tasks to be operated at the edge, i.e., in battery-powered portable devices and other limited-energy environments. In this scenario, scalable resistive memories have been proposed as artificial synapses thanks to their scalability, reconfigurability, and high-energy efficiency, and thanks to the ability to perform analog computation by physical laws in hardware. In this work, we study the material, device, and architecture aspects of resistive switching memory (RRAM) devices for implementing a 2-layer neural network for pattern recognition. First, various RRAM processes are screened in view of the device window, analog storage, and reliability. Then, synaptic weights are stored with 5-level precision in a 4 kbit array of RRAM devices to classify the Modified National Institute of Standards and Technology (MNIST) dataset. Finally, classification performance of a 2-layer neural network is tested before and after an annealing experiment by using experimental values of conductance stored into the array, and a simulation-based analysis of inference accuracy for arrays of increasing size is presented. Our work supports material-based development of RRAM synapses for novel neural networks with high accuracy and low-power consumption.

Resistive switching memory (RRAM) is a promising technology for embedded memory and its application in computing. In particular, RRAM arrays can provide a convenient primitive for matrix-vector multiplication (MVM) with strong impact on the acceleration of neural networks for artificial intelligence (AI). At the same time, RRAM is affected by intrinsic conductance variations, which might cause degradation of accuracy in AI inference hardware. This work provides a detailed study of the multilevel-cell (MLC) programming of RRAM for neural network applications. We compare three MLC programming schemes and discuss their variations in terms of the different slopes in the programming characteristics. We test the accuracy of a two-layer fully connected neural network (FC-NN) as a function of the MLC scheme, the number of weight levels, and the weight mapping configuration. We find a tradeoff between the FC-NN accuracy, size, and current consumption. This work highlights the importance of a holistic approach to AI accelerators encompassing the device properties, the overall circuit performance, and the AI application specifications.

In this letter, we propose an effective route to reduce the cell-to-cell variability in 1T-1R-based random access memories (RRAM) arrays by combining the excellent switching performance of Hf _1-x Al _x O _y with an optimized incremental step pulse with verify algorithm for programming. The strongly reduced cell-to-cell variability improves the thermal and post-programming stability of the arrays, which is relevant for many applications of the RRAM technology. Finally, the retention study at 150 °C enables the prediction of the data storage capability.

Achieving a reliable multilevel programming operation in resistive random access memory (RRAM) arrays is still a challenging task. In this work, we assessed the impact of the voltage step value used by the programming algorithm on the device-to-device (DTD) variability of the current distributions of four conductive levels and on the energy consumption featured by programming 4-kbit HfO ₂ -based RRAM arrays. Two different write-verify algorithms were considered and compared, namely, the incremental gate voltage with verify algorithm (IGVVA) and the incremental step pulse with verify algorithm (ISPVA). By using the IGVVA, a main tradeoff has to be considered since reducing the voltage step leads to a smaller DTD variability at the cost of a strong increase in energy consumption. Although the ISPVA cannot reduce the DTD variability as much as the IGVVA, its voltage step can be decreased in order to reduce the energy consumption with almost no impact on the DTD variability. Therefore, the final decision on which algorithm to employ should be based on the specific application targeted for the RRAM array.