Yeong Hwan Oh 오영환 IRIS Lab, Sungkyunkwan Univ.
Index / Works / 01
Exhibit 01 of 04 Published, Poster
ISLPED ’26 HyperSPACE

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

Making hyperdimensional encoding fully compatible with sparse adder trees on digital compute-in-memory arrays.

Fig. 01 Schematic
HyperSPACE: XOR-binding versus AND-binding Schematic. With XOR-binding about half of the bound bits are ones, so popcount needs an exact adder tree. With AND-binding and sparsity-tuned projection almost every bit is zero, so a sparse adder tree only sums the few non-zeros. FIG. 01 / BINDING SPARSITY FIG. 01 SUBJECT HYPERSPACE SCALE N.T.S. A / XOR-BINDING H K H ⊕ K SPARSITY ρ ≈ 0.5 EXACT ADDER TREE EVERY BIT IS SUMMED + + + + + + + B / AND-BINDING H K H ∧ K SPARSITY ρ > 0.99 SPARSE ADDER TREE ONLY NON-ZEROS ARE SUMMED + + + + + + + ENCODING ENERGY ↓ UP TO 13.03× / ACCURACY ↑ UP TO 3.82%
Fig. 01 XOR-binding pins bound-vector sparsity near 0.5, so popcount must run on exact adder trees. AND-binding compounds sparsity past 0.99, which lets the same computation run on sparse adder trees.

Specifications

As reported in the paper
Encoding energy up to 13.03× lower
Classification accuracy up to +3.82%
Bound-vector sparsity > 0.99
Popcount share of encoding energy (exact adder trees) up to 89.6%
Benchmarks 4

Abstract

ISLPED ’26

An algorithm–hardware co-design that makes HDC encoding fully compatible with sparse adder trees (SATs) on digital SRAM-CIM arrays. Replacing XOR-binding with AND-binding compounds sparsity; sparsity-tuned projection vectors push bound vector sparsity above 0.99; adaptive binary quantization compensates for the non-uniform output statistics that AND-binding induces. 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.

† Co-corresponding authors: K. E. Jeon and J. H. Ko.

Cite

@inproceedings{oh2026hyperspace,
  title = {{HyperSPACE}: Sparse-Adder-Compatible Encoding for Efficient Hyperdimensional Computing on Digital {CIM} Arrays},
  author = {Oh, Yeong Hwan and Kang, Do Yeong and Park, Juhong and Hwang, Chanwook and Jeon, Kang Eun and Ko, Jong Hwan},
  booktitle = {Proceedings of the ACM/IEEE International Symposium on Low Power Electronics and Design (ISLPED '26)},
  location = {Evanston, IL, USA},
  year = {2026},
  doi = {10.1145/3816440.3818512},
  keywords = {Hyperdimensional Computing, In-Memory Computing, Sparse Adder Trees, Approximate Computing}
}