EUVIMEDEUVIMED entdecken
Crossref

S-ANNS: Energy-Efficient Approximate Nearest Neighbor Search via Shift-Based PIM Architecture

Seulki Kim, Yewon Jang, Minho Kim, Daehoon Kim

IEEE Computer Architecture Letters · 2026-07 · 25 (2) · 436-439

DOI: 10.1109/lca.2026.3743912 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Computer Architecture Letters, 2026-07. Verfasst von Seulki Kim und 3 weiteren Autor:innen. Die Arbeit nennt 9 Literaturangaben. Gefördert von: Institute of Information & Communications Technology Planning & Evaluation, Korean government, National Research Foundation of Korea.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Approximate Nearest Neighbor Search (ANNS) underpins Retrieval-Augmented Generation, recommendation, and vector databases but is memory-bandwidth bound at scale, motivating Processing-in-Memory (PIM). Existing PIM architectures, however, trade programmability for efficiency. DPU-based designs support flexible compression, but their general-purpose cores consume DRAM capacity critical to billion-scale corpora, while MAC-based bank-level designs stream only scalar-quantized vectors and cannot support LUT-based compression schemes commonly used for float32 corpora. We present S-ANNS (Shift-ANNS), a bank-level PIM architecture that uses Power-of-Two (PoT) quantization to encode float32 vectors as compact signed-shift representations in DRAM. This representation replaces floating-point multiplication with shift-and-add operations during distance computation while reducing storage overhead. With $S$ S signed terms per dimension, the shift PE reduces per-dimension area and power relative to an FP32 MAC by 49.4% and 58.9% at $S=2$ S = 2 and by 21.6% and 29.2% at $S=4$ S = 4 . At $S=4$ S = 4 , the maximum recall loss is 0.7 percentage points relative to FP32. Across GIST1M, MSTuring1B, DEEP1B, and T2I-100M at recall@10 of 0.9, S-ANNS achieves 1.4–1.9× higher search throughput at 0.3–0.7× the footprint.

Förderung

  • Institute of Information & Communications Technology Planning & Evaluation
  • Korean government
  • National Research Foundation of Korea

Zitieren

APA

Kim, S., Jang, Y., Kim, M., & Kim, D. (2026). S-ANNS: Energy-Efficient Approximate Nearest Neighbor Search via Shift-Based PIM Architecture. IEEE Computer Architecture Letters, 25(2), 436-439. https://doi.org/10.1109/lca.2026.3743912

Vancouver

Kim S, Jang Y, Kim M, Kim D. S-ANNS: Energy-Efficient Approximate Nearest Neighbor Search via Shift-Based PIM Architecture. IEEE Computer Architecture Letters. 2026;25(2):436-439. doi:10.1109/lca.2026.3743912

BibTeX
@article{Kim2026,
  author = {Kim, Seulki and Jang, Yewon and Kim, Minho and Kim, Daehoon},
  title = {S-ANNS: Energy-Efficient Approximate Nearest Neighbor Search via Shift-Based PIM Architecture},
  journal = {IEEE Computer Architecture Letters},
  year = {2026},
  volume = {25},
  number = {2},
  pages = {436--439},
  doi = {10.1109/lca.2026.3743912}
}
Kostenlos starten →