Yeong Hwan Oh 오영환 IRIS Lab, Sungkyunkwan Univ.
Index / Works / 03
Exhibit 03 of 04 Published
DATE ’25 MEMHD

MEMHD: Memory-Efficient Multi-Centroid Hyperdimensional Computing for Fully-Utilized In-Memory Computing Architectures

Replacing one vector per class with many, so hyperdimensional inference fills an in-memory computing array and finishes in a single cycle.

Fig. 03 Schematic
MEMHD: array utilization with one versus many centroids per class Schematic. With one vector per class only a few columns of the IMC array are used, and vectors longer than the array need extra cycles. With multi-centroid class memories every column is used and inference takes one cycle. FIG. 03 / ARRAY UTILIZATION FIG. 03 SUBJECT MEMHD SCALE N.T.S. A / ONE VECTOR PER CLASS A B C D D > ROWS → EXTRA CYCLES IDLE 50% USED, ≥ 2 CYCLES B / MULTI-CENTROID (MEMHD) A1 A2 B1 B2 C1 C2 D1 D2 100% USED, 1 CYCLE ACC. ↑ UP TO 13.69% AT EQUAL MEMORY / MEMORY EFF. ↑ 13.25×
Fig. 03 One class vector per column leaves most of an IMC array idle, and vectors longer than the array need several cycles. Multi-centroid class memories fit the array exactly, so it is fully used and inference takes a single cycle.

Specifications

As reported in the paper
Accuracy at equal memory up to +13.69%
Memory efficiency up to 13.25×
Inference Single cycle, fully utilized array

Abstract

DATE ’25

A memory-efficient multi-centroid hyperdimensional computing framework that replaces single-vector-per-class with multiple class vectors, achieving full IMC array utilization and enabling single-cycle inference through clustering-based initialization and quantization-aware training.

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

Cite

@inproceedings{kang2025memhd,
  title = {{MEMHD}: Memory-Efficient Multi-Centroid Hyperdimensional Computing for Fully-Utilized In-Memory Computing Architectures},
  author = {Kang, Do Yeong and Oh, Yeong Hwan and Hwang, Chanwook and Kim, Jinhee and Jeon, Kang Eun and Ko, Jong Hwan},
  booktitle = {Proceedings of the 2025 Design, Automation \& Test in Europe Conference \& Exhibition (DATE)},
  location = {Lyon, France},
  year = {2025},
  pages = {1--7},
  doi = {10.23919/DATE64628.2025.10993253},
}