記憶する場所で、計算する。
SRAM-CIMの回路とデータパスを設計し、メモリに保存した重みを使って演算を進める仕組みを探ります。
CIRCUIT & ARCHITECTURECOMPUTE-IN-MEMORY · EDGE INTELLIGENCE
Intelligence, closer to memory.
劉昕祐,國立臺灣大學資訊工程學系博士生。研究聚焦於記憶體架構與系統軟體如何協同設計,讓邊緣裝置在有限的能源與運算資源下,可靠地感知、推論並回應真實世界。
01 / RESEARCH FOCUS
エッジでAIを動かすには、演算そのものに加え、データの移動や待ち時間にも目を向ける必要があります。ハードウェアと実行方式の両面から、この課題に取り組みます。
SRAM-CIMの回路とデータパスを設計し、メモリに保存した重みを使って演算を進める仕組みを探ります。
CIRCUIT & ARCHITECTURE依存関係、データの再利用、読み書きの競合を考慮し、複数のCIMコアが効率よく働く実行順序を考えます。
MAPPING & SCHEDULINGEEGNetなどの信号処理を題材に、量子化からRTL検証、FPGAインターフェースまでをつなぎます。
EDGE AI & IMPLEMENTATIONTHE IDEA, MADE VISIBLE
CIMの考え方を、3つの視点から。
下の項目を選んで図を切り替えられます。
メモリアレイを演算にも利用することで、重みの移動を抑える構成を考えます。周辺回路や入出力の設計も重要です。
02 / SELECTED PROJECTS
3件のプロジェクトを表示
2020.02 — 2021.01
在 Pi-Cheng Hsiu 教授指導下,研究能源不穩定時的推論進度保存與恢復。結合 RTOS、TI MSP-EXP430、任務式執行及檢查點處理,維持非揮發性記憶體中的狀態一致性。
RTOS · TI MSP-EXP430 · Checkpoints
2019.02 — 2020.01
在 Yi-Jung Chen 教授指導下,建立週期精確的全系統模擬工具鏈,整合 DDR3/DDR4 時序與 TSV 模型,分析多核心配置、排程、系統布局及功耗。
Sniper · gem5 · SCALE-Sim · DRAMSim · NVSim · McPAT · SimPoint
03 / 學術著作
依發表時間,由新至舊
IEEE Transactions on Circuits and Systems II: Express Briefs
@article{WuPan26,
author = {Wu, Tsung-Yen and Pan, Yi-Shuan and Chen, Yung-Chin and Tseng, Wei-Chun and Lin, Jeng-Fu and Liang, Han-Chung and Yang, Hsin-Che and Yao, Chun-Yen and Wu, Bing-Chen and Chen, Yu-Kai and Chou, Jyun-Jhe and Liu, Xin-You and Wei, Jhen-Yu and Dai, Wei-You and Chen, Chih-Wei and Huang, Sheng-Po and Yuh, Perng-Fei and Wang, Yih and Shih, Chi-Sheng and Liu, Tsung-Te},
title = {N2N-BP2CIM: A 28nm End-to-End Bit-Parallel, Bit-Parallel Computing-In-Memory DNN Processor with High-bandwidth 14T-SRAM Bitcell},
journal = {IEEE Transactions on Circuits and Systems II: Express Briefs},
year = {2026},
month = {oct},
doi = {10.1109/TCSII.2026.3740496},
note = {Early Access}
}International Conference on Compilers, Architecture, and Synthesis for Embedded Systems (CASES’25)
Taipei, Taiwan · 21–22
@inproceedings{LiuShih25,
author = {Liu, Xin-You and Shih, Chi-Sheng and Liu, Tsung-Te and Tsung, Pei-Kuei and Chen, Chih-Wei and Teng, Chieh-Fang},
title = {Device Noises Resilient Training and Inference Framework for Smart Sensing on Analog Computing In Memory},
booktitle = {International Conference on Compilers, Architecture, and Synthesis for Embedded Systems (CASES’25)},
year = {2025},
month = {oct},
pages = {21--22},
location = {Taipei, Taiwan}
}The 51st IEEE European Solid-State Electronics Research Conference
Munich, Germany · 325–328
@inproceedings{KuanYang26,
author = {Pan, Yi-Hsuan and Liu, Hsin-Yu and Huang, Ping-Hsuan and Chen, Wei-Cheng and Yu, Yue-Long and Chou, Jyun-Jhe and Liu, Xin-You and Teng, Chieh-Fang and Chen, Chih-Wei and Chung, Pei-Kuei and Shih, Chi-Sheng and Liu, Tsung-Te},
title = {An LLM-Friendly CIM-Based Transformer Accelerator with Multi-Head Adaptable Computation Scheme},
booktitle = {The 51st IEEE European Solid-State Electronics Research Conference},
year = {2025},
month = {sep},
pages = {325--328},
location = {Munich, Germany},
doi = {10.1109/ESSERC66193.2025.11214103}
}The 39th ACM/SIGAPP Symposium On Applied Computing
Avila, Spain
@inproceedings{Liu24,
author = {Liu, Xin-You},
title = {Real-Time 3D Registration and Fusion with SRAM-based Analog In-Memory Computing},
booktitle = {The 39th ACM/SIGAPP Symposium On Applied Computing},
year = {2024},
month = {apr},
location = {Avila, Spain},
publisher = {ACM SIGAPP},
language = {English},
doi = {10.1145/3605098.3635170}
}Sensors 22(9), 3491
@article{Chou2022,
author = {Chou, Jyun-Jhe and Chang, Ting-Wei and Liu, Xin-You and Wu, Tsung-Yen and Chen, Yu-Kai and Hsu, Ying-Tuan and Chen, Chih-Wei and Liu, Tsung-Te and Shih, Chi-Sheng},
title = {CIM-Based Smart Pose Detection Sensors},
journal = {Sensors},
year = {2022},
month = {may},
volume = {22},
number = {9},
eid = {3491},
doi = {10.3390/s22093491}
}04 / BEHIND THE RESEARCH
PhD Student · Computer Science & Information Engineering
National Taiwan University
目前於國立臺灣大學資訊工程學系攻讀博士,研究計算機架構、即時系統與記憶體內運算。在台灣半導體研究中心(TSRI)的研究,探索以 SRAM/MRAM CIM 支援智慧感測裝置上的 EEGNet 推論,並考量容錯方法與執行效率。聯發科臺大研究中心與中央研究院的研究經驗,則讓我進一步串連系統軟體、邊緣 AI 與間歇式運算,思考如何將模型與硬體協同設計,打造可靠且節能的智慧系統。
05 / 學歷與經歷
— 至今
國立臺灣大學
—
國立臺灣大學
—
國立暨南國際大學
畢業班級排名 1/52
— 至今
博士生 研究助理
研究智慧感測裝置上的 EEGNet 記憶體內運算,探索 SRAM 與 MRAM CIM 架構,並設計非揮發性記憶體系統的容錯演算法。透過分析 EEGNet 執行軌跡,評估適合高效能推論的系統設計。
— 至今
系統軟體工程師
研究行動裝置上的邊緣語言模型、影像去雜訊與生成模型,分析快速 token 生成所需的加速器設計,並以貝氏機率模型開發高速校正演算法,降低轉換誤差。
—
大學部 研究助理
研究間歇式供電裝置上的視覺神經網路,利用非揮發性記憶體保存執行進度,並比較任務式與檢查點式恢復機制。
06 / 獎項與肯定
07 / 教學、交流與認證
2024 秋季
2022 春季
CIM for Transformer Network and LLM
2025.03
Real-Time 3D Registration with Defect Knowledge Transfer on SRAM Analog In-Memory Computing
2024
Exploring the Accelerator Design in the Future: Robust In-memory Computing and Searching Architecture for 3D Computer Vision
2024.06
Low-Power Flexible-Bit CIM for Always-On Real-Time 3D Sensing
2023.10
Image Super-Resolution Using Convolutional Networks for Human Pose Detection
Better systems begin with
closer connections.