A 22nm 41.8TFLOPS/W AI-Edge Transformer/CNN Nonvolatile-Processor Using QKV-Softmax-Layer-Fused Hybrid ReRAM-CIM and Concurrent-Transpose/Non-Transpose SRAM-CIM

Tiny AI-edge devices use nvCIM for power-off weight storage and active-mode computation, enabling high energy efficiency (EF) and low power-on latency. While tiny Transformer models offer higher accuracy than CNNs, they pose significant challenges for homogeneous nvCIM architectures due to attention-based operations that require dynamic matrix multiplications (MM), where both inputs and weights are generated on-the-fly. This paper presents a hybrid ReRAM-CIM and SRAM-CIM nvProcessor optimized for Transformer and CNN models, featuring a CIM-friendly layerfused flow. Fabricated in a 22 nm process, the proposed nvProcessor achieves 41.8 TFLOPS/W (system), 55.2 TFLOPS/W (macro) and 74.48% ImageNet accuracy, enabling high EF and accuracy with tiny (Mb-level) models.

Authors

Institutions

Publication Details

Published
2025-06-08
DOI
https://doi.org/10.23919/vlsitechnologyandcir65189.2025.11074985
Citations
1
Primary Topic
Advanced Memory and Neural Computing
Type
article
Field-Weighted Citation Impact
0.73
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
article

A 22nm 41.8TFLOPS/W AI-Edge Transformer/CNN Nonvolatile-Processor Using QKV-Softmax-Layer-Fused Hybrid ReRAM-CIM and Concurrent-Transpose/Non-Transpose SRAM-CIM

Ren-Shuo Liu, Hung-Hsi Hsu, Meng‐Fan Chang, Chung‐Chuan Lo et al.
1 citations
Advanced Memory and Neural Computing
0.73
article

A 22nm 41.8TFLOPS/W AI-Edge Transformer/CNN Nonvolatile-Processor Using QKV-Softmax-Layer-Fused Hybrid ReRAM-CIM and Concurrent-Transpose/Non-Transpose SRAM-CIM

Ren-Shuo Liu, Hung-Hsi Hsu, Meng‐Fan Chang, Chung‐Chuan Lo, Kea‐Tiong Tang, Chih-Cheng Hsieh, Shih-Hsin Teng, Wei‐Ting Hsu, Chang-Yuan Chen, Ping-Sheng Wu, Mon‐Shu Ho, Win-San Khwa, Tai-Hao Wen, Yen-Che Huang, H. L. Lu, Yu‐Chen Chang, Hua-Jin Wen, Yu-Der Chih, C.C. Wu, Tsung-Yung Jonathan Chang
article en
1 citations

Abstract

Tiny AI-edge devices use nvCIM for power-off weight storage and active-mode computation, enabling high energy efficiency (EF) and low power-on latency. While tiny Transformer models offer higher accuracy than CNNs, they pose significant challenges for homogeneous nvCIM architectures due to attention-based operations that require dynamic matrix multiplications (MM), where both inputs and weights are generated on-the-fly. This paper presents a hybrid ReRAM-CIM and SRAM-CIM nvProcessor optimized for Transformer and CNN models, featuring a CIM-friendly layerfused flow. Fabricated in a 22 nm process, the proposed nvProcessor achieves 41.8 TFLOPS/W (system), 55.2 TFLOPS/W (macro) and 74.48% ImageNet accuracy, enabling high EF and accuracy with tiny (Mb-level) models.

National Chung Hsing University (TW), National Tsing Hua University (TW), Taiwan Semiconductor Manufacturing Company (Taiwan) (TW)
Affordable and clean energy
Openalex Percentile: Top 25%
Advanced Memory and Neural Computing
0.73
DOISourceOpenAlex
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.