Search

Memristors, Flash, and related nonvolatile analog device technologies offer in‐memory computing structures operating in the analog domain, such as accelerating linear matrix operations in array structures. These take advantage of analog tunability and large dynamic range. At the other side, content addressable memories (CAM) are fast digital lookup tables which effectively perform nonlinear Boolean logic and return a digital match/mismatch value. Recently, nonvolatile analog CAMs have been presented merging analog storage and analog search operations with digital match/mismatch output. However, CAM blocks cannot easily be inserted within a larger adaptive system due to the challenges of training and learning with binary outputs. Here, a missing link between analog crossbar arrays and CAMs, namely a differentiable content addressable memory (dCAM), is presented. Utilizing nonvolatile memories that act as a "soft" memory with analog outputs, dCAM enables learning and fine‐tuning of the memory operation and performance. Four applications are quantitatively evaluated to highlight the capabilities: improved data pattern storage, improved robustness to noise and variability, reduced energy and latency performance, and an application to solving Boolean satisfiability optimization problems. The use of dCAM is envisioned as a core building block of fully differentiable computing systems employing multiple types of analog compute operations and memories. [ABSTRACT FROM AUTHOR]

Nanoscale memristors in a crossbar configuration have demonstrated their ability to accelerate modern computing workloads in various applications, including machine learning, image processing, and data analytics. Modeling the crossbar behavior is critical to software-hardware co-design, but most previous works focused on single or several memristor devices. So, challenges still exist in large-scale crossbar implementations due to non-idealities that only emerge at the system level. In this brief, we build a crossbar model based on experimentally characterized device statistics in large crossbar arrays. We identify different types of imperfections, including statistical device relaxation, fluctuation, peripheral circuits, etc. The experimentally-validated model is then used to co-optimize analog matrix multiplication and neural network applications. Specifically, we propose and implement defect-aware training and verify that the neural network trained with our algorithm can provide better accuracy and reliability when deployed to physical crossbars. Finally, we achieve an experimental accuracy of 93.4% on the MNIST dataset in a physical crossbar by training based on our crossbar model to compensate for statistical stochasticity, 8.4% higher than the vanilla model. More importantly, the accuracy remains larger than 90.0% after two days, while the accuracy with the vanilla model drops to 73.2% because of conductance relaxation. The method is also scaleable to more practical networks and demonstrates a 92.3% CIFAR-10 accuracy with the VGG-16 model on the simulated crossbar model. [ABSTRACT FROM AUTHOR]