Computing-In-Memory (CIM) is widely applied in neural networks due to its unique capability to perform multiply-and-accumulate operations within a circuit array. This process directly obtains the current value through...
详细信息
Computing-In-Memory (CIM) is widely applied in neural networks due to its unique capability to perform multiply-and-accumulate operations within a circuit array. This process directly obtains the current value through the product of voltage and conductance, accumulating it on the bit line, thus realizing storage and computing functionalities simultaneously within a single array. This significantly reduces the power consumption and time delay in data processing. Unfortunately, implementing general-purpose logic computations in large-scale memory arrays with CIM remains a challenge. This paper introduced a novel device concept, the programmable diode-a special type of memristor with a high switching window, ideally suited for memory arrays to reduce power consumption. A compact SPICE model was developed to enable circuit-level simulations in EDA tools. We also proposed a method to efficiently control the programmable diode for logic operations in memory arrays, and in this way, we constructed a parallel 8-bit full adder to verify the feasibility of the proposed method. Finally, based on the 8-bit full adder, we built a 5kb in-memory logicarray capable of executing logic computations and simulated it using EDA tools. The simulation results demonstrated that the 5kb in-memory logicarray can perform fundamental Boolean logic and arithmetic operations with high repeatability and parallelism, perfectly realizing the functionality of in-memory logic computation. Our work can provide a feasible scheme for realizing large-scale general logic computation systems based on CIM.
暂无评论