[C101] HyperSPACE: Sparse-Adder-Compatible Encoding for Efficient Hyperdimensional Computing on Digital CIM Arrays

Abstract

Hyperdimensional computing (HDC) is attractive for edge AI, but in digital SRAM-CIM implementations its encoding stage is energy-dominant, with popcount in exact adder trees (EATs) accounting for up to 89.6% of encoding energy. Sparse adder trees (SATs) can cut this cost by nearly 10x, yet standard XOR-based binding fixes vector sparsity near 0.5 and therefore cannot exploit SATs. We propose HyperSPACE, an algorithm-hardware co-design that makes encoding fully SAT-compatible. HyperSPACE replaces XOR-binding with AND-binding to compound sparsity, pairs it with sparsity-tuned projection vectors to push vector sparsity beyond 0.99, and introduces adaptive binary quantization to compensate for the non-uniform output statistics that AND-binding induces. Evaluated on 128x64 SRAM-CIM arrays, HyperSPACE reduces encoding energy by up to 13.03x while improving classification accuracy by up to 3.82% across four benchmarks.