2T2R RRAM-Based In-Memory Hyperdimensional Computing Encoder for Spatio-Temporal Signal Processing

Hyperdimensional computing (HDC) is a brain-inspired computational framework that exploits hypervectors as an alternative to computing with numbers. In-memory computing implementation of HDC (IM-HDC) provides a robust and energy-efficient approach to process spatio-temporal (ST) signals since it significantly reduces data transfer overhead. However, previous IM-HDC suffers from the large peripheral circuit overheads to assist the component-wise hypervector operations. To address these issues, we propose a voltage-mode two-transistor-two-resistor (2T2R) RRAM-based IM-HDC encoder for ST signal processing. The in-memory hyperdimensional encoding has been achieved by performing binding/bundling in the 2T2R RRAM array and permutation on specially designed digital peripheral circuits. Combining with an RRAM-based in-memory associative search module, we validated an average classification accuracy of 97.96% on gesture recognition of electromyogram (EMG) signals, while achieving high robustness and throughput, low latency, and 39× higher energy-efficiency as compared to current state-of-the-art IM-HDC encoders.