DropHD: Technology/Algorithm Co-Design for Reliable Energy-Efficient NVM-Based Hyper-Dimensional Computing Under Voltage Scaling

Brain-inspired hyperdimensional computing (HDC) offers much more efficient computing compared to other classical deep learning and related machine learning algorithms. Unlike classical CMOS, emerging non-volatile memories (NVMs) used in the realization of HDC are susceptible to failures under voltage scaling, which is essential for energy saving. Although HDC is inherently robust against errors, this is only possible when hypervectors with a large dimension (e.g., 10,000 bits) are being used, resulting in significant energy consumption. This work demonstrates, for the first time, that different NVM technologies exhibit different error characteristics under voltage scaling. In contrast to conventional CMOS-based SRAM, we demonstrate that the error behavior is data-dependent and not captured by simple bit flips in emerging NVMs. We employ our cross-layer framework that starts from the underlying technology all the way up to the algorithm to develop the novel HDC training approach DropHD. DropHD considerably shrinks the size of hypervectors (e.g., from 10,000 bits down to merely 3000 bits), while maintaining a high inference accuracy. The use of aggressive voltage scaling reduces energy consumption by 1.6 x. DropHD further reduces it to up to 9.5 × while fully recovering the induced accuracy drop, i.e. without a tradeoff.