EUVIMEDEUVIMED entdecken
Crossref

A 22 nm Floating-Point ReRAM Compute-in-Memory Macro Using Residue-Shared ADC for AI Edge Device

Hung-Hsi Hsu, Tai-Hao Wen, Win-San Khwa, Wei-Hsing Huang, Zhao-En Ke, Yu-Hsiang Chin, Hua-Jin Wen, Yu-Chen Chang, Wei-Ting Hsu, Ashwin Sanjay Lele, Bo Zhang, Ping-Sheng Wu und 9 weitere

IEEE Journal of Solid-State Circuits · 2025-01 · 60 (1) · 171-183

DOI: 10.1109/jssc.2024.3470211 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Journal of Solid-State Circuits, 2025-01. Verfasst von Hung-Hsi Hsu und 20 weiteren Autor:innen. Die Arbeit wurde bisher 23-mal zitiert (Crossref) und nennt 37 Literaturangaben. Gefördert von: Taiwan Semiconductor Manufacturing Company (TSMC)-Joint Developed Project, National Science and Technology Council (NSTC) of Taiwan.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Artificial intelligence (AI) edge devices increasingly require the enhanced accuracy of floating-point (FP) multiply-and-accumulate (MAC) operations as well as nonvolatile on-chip memory to minimize the movement of weight data in power-off mode. Designing non-volatile compute-in-memory (nvCIM) macros for FP operations imposes several challenges, including: 1) a tradeoff between inference accuracy and weight bit-width following pre-alignment; 2) long computing latency and high energy consumption; 3) large cell array current during computation; and 4) high multi-bit readout energy consumption. In this study, we devised four schemes to address these issues, including: 1) a kernel-wise weight pre-alignment (K-WPA); 2) a rescheduled multi-bit input compression (RS-MIC); 3) HRS-favored dual-sign-bit (HF-DSB); and 4) residue-shared analog-to-digital converter (RS-ADC). A 16 Mb resistive random access memory (ReRAM) nvCIM macro fabricated for FP operations using foundry-provided ReRAM (22 nm CMOS technology) achieved an efficiency of 34.2 TFLOPS/W under BF16-input, BF16-weight, and FP32-output and 31.4 TFLOPS/W under FP16-input, FP16-weight, and FP32-output.

Förderung

  • Taiwan Semiconductor Manufacturing Company (TSMC)-Joint Developed Project
  • National Science and Technology Council (NSTC) of Taiwan

Zitieren

APA

Hsu, H., Wen, T., Khwa, W., Huang, W., Ke, Z., Chin, Y., Wen, H., Chang, Y., Hsu, W., Lele, A. S., Zhang, B., Wu, P., Lo, C., Liu, R., Hsieh, C., Tang, K., Teng, S., Chou, C., Chih, Y., …, & Chang, M. (2025). A 22 nm Floating-Point ReRAM Compute-in-Memory Macro Using Residue-Shared ADC for AI Edge Device. IEEE Journal of Solid-State Circuits, 60(1), 171-183. https://doi.org/10.1109/jssc.2024.3470211

Vancouver

Hsu H, Wen T, Khwa W, Huang W, Ke Z, Chin Y, et al. A 22 nm Floating-Point ReRAM Compute-in-Memory Macro Using Residue-Shared ADC for AI Edge Device. IEEE Journal of Solid-State Circuits. 2025;60(1):171-183. doi:10.1109/jssc.2024.3470211

BibTeX
@article{Hsu2025,
  author = {Hsu, Hung-Hsi and Wen, Tai-Hao and Khwa, Win-San and Huang, Wei-Hsing and Ke, Zhao-En and Chin, Yu-Hsiang and Wen, Hua-Jin and Chang, Yu-Chen and Hsu, Wei-Ting and Lele, Ashwin Sanjay and Zhang, Bo and Wu, Ping-Sheng and Lo, Chung-Chuan and Liu, Ren-Shuo and Hsieh, Chih-Cheng and Tang, Kea-Tiong and Teng, Shih-Hsin and Chou, Chung-Cheng and Chih, Yu-Der and Chang, Tsung-Yung Jonathan and Chang, Meng-Fan},
  title = {A 22 nm Floating-Point ReRAM Compute-in-Memory Macro Using Residue-Shared ADC for AI Edge Device},
  journal = {IEEE Journal of Solid-State Circuits},
  year = {2025},
  volume = {60},
  number = {1},
  pages = {171--183},
  doi = {10.1109/jssc.2024.3470211}
}
Kostenlos starten →