AI) exploits large datasets to learn how to solve a broad variety of tasks such as text/ image recognition, data clustering, equity pricing, and many others [3] . However, computers based on the Von-Neumann architecture are inefficient in the execution of modern AI workloads, and the required power becomes unsustainable due to the growing volume of data involved.The bottleneck for performance derives from the memory wall. The bandwidth of processors has grown faster than the memory one, with the difference between them diverging exponentially. [4] The energetic inefficiency originates to a large extent from energy-hungry accessing of off-chip memory. [5] With the end of Moore's law approaching [6] and the rise of AI, the quest for alternative computing paradigms has intensified.In-memory computing has emerged as a promising hardware architecture for data-centric applications. It consists of memory cells interconnected in a specific design to locally execute logic or arithmetical operations. [7] The crossbar array is an example of an in-memory computing architecture that raised a strong interest in the electron devices community. In this architecture, the memory elements are positioned at the crosspoint between bitlines and wordlines, [8] as depicted in Figure 1a. When a vector of read voltages (V) is applied on the array elements, the currents (I) establishing at The in-memory computing paradigm aims at overcoming the intrinsic inefficiencies of Von-Neumann computers by reducing the data-transport per arithmetic operation. Crossbar arrays of multilevel memristive devices enable efficient calculations of matrix-vector-multiplications, an operation extensively called on in artificial intelligence (AI) tasks. Resistive random-access memories (ReRAMs) are promising candidate devices for such applications. However, they generally exhibit large stochasticity and device-to-device variability. The integration of a sub-stoichiometric metal-oxide within the ReRAM stack can improve the resistive switching graduality and stochasticity. To this purpose, a conductive TaO x layer is developed and stacked on HfO 2 between TiN electrodes, to create a complementary metal-oxide-semiconductorcompatible ReRAM structure. This device shows accumulative conductance updates in both directions, as required for training neural networks. Moreover, by reducing the TaO x thickness and by increasing its resistivity, the device resistive states increase, as required for reduced power consumption. An electric field-driven TaO x oxidation/reduction is responsible for the ReRAM switching. To demonstrate the potential of the optimized TaO x /HfO 2 devices, the training of a fully-connected neural network on the Modified National Institute of Standards and Technology database dataset is simulated and benchmarked against a full precision digital implementation.