BMX-FPCA: 3D Beyond-Moore Flexible Field Programmable Crossbar Array Architecture

Hasita Veluri and Dilip Vasudevan
Lawrence Berkeley National Laboratory


An FPCA (Field-Programmable Crossbar Array) architecture based on memristors offers improved programmability while blurring the line between computing and memory. However, the high-power analog-digital interface and device irregularities create performance bottlenecks. Here, we propose a unique memory-centric, reconfigurable, general-purpose and emerging devices flexible 3D in-memory computing platform: BMX-FPCA, capable of executing analog, digital, and memory tasks while managing massive amounts of data effectively. We accomplish maximum energy efficiency through the interface, crossbar layout optimization, and dynamic distribution of the available resources for memory, arithmetic, and analog computing operations. Further, the proposed architecture achieved a throughput improvement of 308.8× for neural network implementation over the state-of-the-art.