Word Line Underdrive Technique to Increase Signal Swing and Linearity for In-Memory Compute Applications (IMC) in 65nm LSTP technology

The movement of data between memory and compute limits the system’s throughput for machine learning applications due to the Von-Neumann bottleneck. In-memory computation (IMC) has enormous potential to overcome Von-Neumann bottleneck for data-intensive applications where massive data movement occurs between memory and compute units. IMC reduces the latency and and increases the energy efficiency of the system. Mixed signal IMC requires multi-wordline activation. This paper presents the wordline (WL) underdrive and current mirroring techniques to 1) eliminate the spurious write problem and 2) enhance the system’s linearity. 20% WL underdrive increases the linearity of the proposed architecture by 36%. At FS/125°C corner to get a full swing on bitline, nearly 50% reduction in wordline voltage is required to prevent the spurious write operation.