Junhyuk Lee

ARM CME · On-device AI · SW/HW Co-design & Evaluation

Junhyuk Lee

ARM CME · On-device AI · SW/HW Co-design & Evaluation

I currently focus on ARM CME, on-device AI software, and SW/HW co-design and evaluation. My experience at Samsung Electronics spans mobile NPU architecture and evaluation, AI frameworks, application processors, and multimedia optimization.

Email GitHub LinkedIn About

Experience

Samsung Electronics

Principal Engineer

Mar 2023–Present

  • Mobile NPU architecture and evaluation, including work with AI frameworks.

Software Engineer

Aug 2019–Present

  • Mobile NPU model optimization and hardware analysis based on mobile use-case requirements.

Senior Software Engineer

Feb 2009–Present

  • Application processor system analysis and benchmarking using C/C++ and Python, including Neural Engine hardware IP and Qualcomm SNPE evaluation.
  • TensorFlow Mobile/Lite performance and power evaluation, on-device personalization, and participation in Samsung Mobile's MLPerf efforts.

Creative Lead

Aug 2018–Jul 2019 · Samsung C-Lab, Photo Teller

  • Led Photo Teller, using CNN, Word2Vec, and FastText, with TensorFlow/Keras for CNN/LSTM text generation on Android.

Technical Business Manager

May 2015–Jun 2017 · Samsung Electronics GmbH, Frankfurt

  • Coordinated with headquarters, supported carriers and B2B customers, and worked on local adaptation and technical workshops.

Software Engineer

Jun 2012–May 2015

  • Application processor solutions and wearable audio, including Yamaha Y831 bring-up.

Software Engineer

Feb 2009–Jun 2012

  • Tizen multimedia, Twitter, and instant messaging applications.

Expertise

  • C/C++
  • Python
  • Linux
  • TensorFlow / TensorFlow Lite
  • Keras
  • NPU architecture and evaluation
  • Audio codec / BSP
  • Arm FVP / AOSP bring-up

Credentials

  • Samsung SW Certification Pro
  • Data Science Level 2 · Sep 2024
  • OPIc AL · Jul 2017

Engineering Work

AirPods Android Companion

Built an Android companion app that separates nearby BLE advertisements from connected-device battery data, verifies noise-control replies over AAP/L2CAP, and coordinates Activity, foreground monitoring, and widget state.

Read the Korean technical article and build the app.

Arm FVP / Android Platform Bring-up

Investigated Android Open Source Project (AOSP) bring-up on Arm Fixed Virtual Platforms (FVP), tracing boot dependencies across kernel and init, APEX and linker namespaces, ART/JNI, framework services, and graphics. Documented startup and application-launch trace variability, including reproducibility factors such as runtime state, Binder activity, rendering, and scheduling.

  • Recorded a successful FVP boot milestone in May 2026; follow-up platform patches were still required.
  • Analyzed service-loading and graphics initialization failures while distinguishing observed progress from planned stabilization work.

Read the detailed FVP work record.

Videos

S26: 1.5B sLM on CPU

A CPU demonstration of a 1.5B small language model on S26. Video: 0:19.

Watch the original on LinkedIn
Flip6: 1.8B sLM on CPU

A CPU demonstration of a 1.8B small language model on Flip6. Video: 1:04.

Watch the original on LinkedIn
PyTorch Mobile CPU demo

A demonstration of PyTorch Mobile running on CPU. Video: 0:15.

Watch the original on LinkedIn

Selected Patents

I have filed 15+ patent applications. Selected publications and granted patents are listed below.

  1. Electronic device and operation method of electronic device for performing calculation using artificial intelligence model US20240202590A1 · Patent application publication
  2. Apparatus and method for determining a learning model for learning KR20240026050A · Patent application publication
  3. Electronic device and method of recognizing a force touch, by electronic device US11899884B2 · Granted patent document
  4. Electronic apparatus and controlling method thereof US11467642B2 · Granted patent document
  5. Electronic device and learning model determination method for learning of electronic device US20240077970A1 · Patent application publication
  6. Electronic device for processing neural network model and method of operating the same US20230072337A1 · Patent application publication
  7. Electronic device identifying force touch and method for operating the same US11874995B2 · Granted patent document
  8. Smart Alarm service using emoticon in instant messenger J2011-07-5800 · LinkedIn reference

BLE 광고와 연결된 기기의 배터리 상태를 구분하고, AAP/L2CAP으로 듣기 모드 응답을 확인하는 Android 앱의 구현과 실기기 검증 범위를 정리했습니다.

Recent Posts

deeplearning Model for android

less than 1 minute read

DeepLearning 관련해서 공부하면서 결과물을 눈으로 보고 확인하고 싶어서 Android App 으로 산출물을 만들었습니다. 최대한 간단히 만든다고 만들었는데 UX 와 실력이 부족해서 아직은 목표대비는 부족 한것 같습니다.

re training sustain

less than 1 minute read

Tensorflow 로 학습을 시키다 보면 HW / 시간의 제약으로 한번 끊고 학습을 다시 시켜야할 경우가 생긴다 이랄떠 save.restore 라는 함수를 사용해서 연속해서 학습을 시키면 된다

Quantization awareness training for Flownet

less than 1 minute read

the flownet is most popuar Neural Network model for differenct view from camera based on multi input. I would like to check out the result of the quantizati...

Welcome to my Webpage !

less than 1 minute read

You’ll find this post in my blog on github _posts directory. I hope to remind my old memories with old friends. I would like to step in first in github web...