AttnACQ: Attentioned-AutoCorrelation-Based Query for Hyperdimensional Associative Memory

The low power and latency of hyperdimensional associative memory (HAM) promotes hyperdimensional computing (HDC)’s efficiency. However, overheads of HAM can be hardly further reduced, since HAM needs to store all class hypervectors, and complete the similarity calculation between them and the input sample hypervector. To address this issue, an attentioned-autocorrelation based query method called AttnACQ is proposed to avoid the storage and operations corresponding to class hypervectors. Furthermore, levearging SOT-MRAM’s ultra-low read/write delay, a high-efficient HAM matching to AttnACQ is proposed. Experiments show that this AttnACQ boosted HAM saves more than 95% area and 50% latency.