STT-MRAM Based Highly Orthogonal Hypervector Generator for Hyperdimensional Computing

As an emerging computing paradigm, hyperdimensional computing (HDC) features low computational cost, fast learning speed, low hardware overhead, high robustness, high independence of data volume and excellent model interpretability. In order to use HDC in learning and classification, the dataset of all symbols need to be randomly mapped into a space of hypervectors firstly. The precision of mapping is mainly determined by the randomness, which will finally affect the efficiency of learning and classification. This paper proposes a highly orthogonal hypervector generator by utilizing the stochastic switching behavior of spin transfer torque magnetic random access memory (STT-MRAM). With true randomness in the switching process, the generation of hypervectors is more efficient. In addition, the design method of co-optimization is used to explore the design space.