A Novel Current Comparator Enabling Large RRAM Crossbars for BNNs and PUFs

Emerging non-volatile memory (NVM) device technologies are advancing in-memory computing (IMC) applications by providing faster computation speeds and reducing resource overhead. Crossbar structures are commonly used in IMC with NVM devices such as memristors to perform matrix-vector multiplication for deep learning and security primitive applications. Large crossbars are required to implement today's deep learning models, especially for implementing specialized Binarized Neural Networks (BNNs) architectures and to construct security primitives such as physical unclonable functions (PUFs). To digitize crossbar currents, current-sense comparators based on current mirrors are widely used in BNN crossbars. However, conventional comparators have a limited current range due to the decrease of the bit-line voltage as the number of active crossbar elements connected to the bit-line increases. This paper presents a current comparator that employs a regulated cascode sensing stage which boosts the input current range by stabilizing the bit-line voltages. By increasing the current range, a larger crossbar size can be supported. Extensive simulations were carried out to verify the proposed technique, demonstrating that a significantly larger crossbar size can be achieved with the proposed comparator compared to a traditional current-sense comparator for the same area and resolution. Designed in a 180 nm technology, the proposed comparator achieves a resolution of 50nA and 100μA for PUF and BNN applications, respectively, and dissipates 198 μW,