About Me
I am a Machine Learning Engineer at Intel. I earned a dual master’s degree from Tongji University (2017-2021) and KIT (2019-2020), focusing on deep learning, human-machine interaction, and driving behavior analysis. I also hold a bachelor’s degree from Hefei University of Technology. My research interests include Machine Learning, Automated Vehicles, and Intelligent Robot/UAV systems
Education
- M.Eng || Tongji University (TJU) || Vehicle Engineering || 09/2017-03/2021
- M.Sci || Karlsruhe Institute of Technology (KIT) || Mechanical Engineering || 04/2019-08/2020
- B.Eng || Hefei University of Technology (HFUT) || Vehicle Engineering || 09/2013-07/2017
Research Experience
DL-Based CPU Performance Prediction Model
Machine Learning Engineer || Project PI
Intel, Shanghai | Sep.2022 - Present
- Overview: Leverage DL models to predict CPU performance on multiple generations of Intel® Xeon® Scalable Processorss. The task involves producing accurate benchmarks, developing deep neural networks, and data processing on CPU performance prediction.
- Highlights:
- Implemented two custom deep learning models (MambaCPU and NCPP) for efficient CPU performance prediction with less than 3% error across a variety of performance benchmarks such as SPEC CPU, Memory Latency Checker, HPCG, and STREAM. NCPP uses a group attention mechanism, while MambaCPU employs a state-space approach.
Contributing to faster issue diagnosis related to microarchitecture, kernel, and memory in post-silicon debugging in cluster data, with potential applications of the proposed model in the CPU design phase through their prediction results. This works also led to two papers.
Want to read more? Click above icons to get more info Black link for NCPP Blue link for Mamba
Maneuver identification in urban scene using machine learning
Master Thesis || Advisor: R. Sci Thilo Braun
KIT | Jan.2020 - Aug.2020
- Overview: Developed a model for classifying urban driving behaviors, contributing to advancements in intelligent transportation systems and smart vehicle technology.
- Highlights:
- Utilized real-world time-series data of vehicle driving behavior (e.g., speed, position) collected via GPS and radar.
- Applied multiple NN architectures to classify driving behaviors in urban environments.
- Addressed key challenges, including class imbalance in driving behavior datasets, high-dimensional features, and capturing temporal dependencies.
- Achieved a final accuracy of 93%, improving by 13% compared to traditional methods. The model was simpler, lighter, and more efficient.
A machine learning based EV battery charging prediction method
Team Leader, PI || Advisor: Prof Su Zhou
TJU | Jul.2018 - Dec.2018
- Overview: The prediction of power battery charging energy is the core issue of power battery decline assessment. Integrated AI/ML methods to perform regression analysis for predicting vehicle charging energy, aimed at improving the estimation of electric vehicle battery life.
- Highlights:
- Conducted extensive data preprocessing and feature engineering, including abnormal correction and missing data imputation.
- Utilized four linear regression models (Lasso and Ridge regression) as well as two ensemble-based XGBoost models as benchmarks.
- Combined these models using a stacking method based on their performance.
- Applied 10-fold cross-validation to improve algorithm stability and generalization under large data scales.
- Achieved a prediction accuracy of 98% for vehicle charging energy
Outcome: National Silver Award (2nd / 200+)
IoT Supply Chain Optimization with Machine Learning for Bosch
Team Leader, PI
TJU | Mar.2018 - Jul.2018
- Overview: Implemented multiple AI-based algorithms to optimize the intelligent application of logistics, specifically focusing on the circular reuse of turnover boxes within supply chains.
- Highlights:
- Leveraged QR codes and RFID to streamline tracking of boxes and trucks, cutting hardware use by 1/1000.
- Integrated environmental data using Microsoft IoT-hub and built a web app for data management.
- Employed K-means for user classification, genetic algorithms for route planning, and scheduling algorithms for box allocation.
- Enabled cost-effective tracking and management of turnover boxes using AI-based algorithms.
- Outcome: National Champion, invited to global Hackathon in Berlin 2019
Work Experience
Intel || Machine Learning Engineer || Full time
Data Center and Artificial Intelligence Group (DCAI), Shanghai
Sep.2022 - Present
- Optimizing AI Model Training Performance on Intel Gaudi-2D
- Responsible for training the LLaMA 3.1-8B model using the Megatron-DeepSpeed framework integrated into Intel’s HabanaAI, deployed on an 8-card Intel Gaudi-2D platform.
- Focusing on evaluating and optimizing the model’s throughput performance to ensure training efficiency and resource utilization.
- Aiming to systematically identify and resolve bottlenecks affecting throughput.
- Collaborate with the technical team to analyze data and provide performance optimization recommendations, further enhancing Intel GPU’s competitiveness in AI training.
- AI Model Inference Optimization on Intel Xeon CPUs:
- Expertly utilized OpenVINO framework to perform inference with AI models (BERT-LARGE, GPT-J-6B, LLaMA-2-7B, ResNet50, Stable-Diffusion, YOLOv5s) on Intel server CPUs. Focused on achieving optimal performance across precision levels such as AVX_fp32, AMX_INT8, AMX_BF16, and AMX_FP16, setting industry benchmarks for Intel CPU AI model performance.
- Developed a fully automated workflow in Docker containers using Python to perform AI inference with OpenVINO on Intel CPUs, which streamlined model execution, data extraction, and database uploads
- Enhancing Debugging Efficiency in Server CPU Clusters: Developed a predictive model to identify problematic cases in benchmark data across server CPU clusters, enhanced the efficiency of debugging engineers and reduced post-silicon debugging cycles.
- Outcome: Data published on Intel’s website, Mamba, NCPP
Baidu || Algorithm Engineer || Full time
Phoenix Nest, Product Strategy R&D Department || Shanghai
Mar. 2021 – Sep.2022
- Deep Learning Ad Ranking: Utilized deep learning methods to construct the user image by searching keywords and historical click frequency. Ads were ranked according to AUC scores, with the most relevant ads prioritized. Integrated users click-session history to improve click-through rates.
- Offline Training and Online Deployment of Big Data Models:
- Managed big data extraction using Parquet and initial feature processing with Hadoop.
- Trained extensive models with Baidu PaddlePaddle, handling sample sizes in the tens of millions.
- Conducted offline validation and A/B testing for strategy optimization.
- Deployed models for real-time online inference.
- Developed and integrated C++ modules as needed.
- Outcome: Resulted in a 6% increase in overall online revenue and a 2.2% improvement in click-through rates
NIO || Intern
Department of Motor Systems || Shanghai
Nov.2018 – Mar.2019
Overview: Responsible for the assembly of electric motors for NIO models ET7, ES6, and ES8, as well as the maintenance of 2D engineering drawings and data.
Utilized CATIA software to assemble motor components and integrate them into the vehicle chassis, generating various views for maintenance and verification, and providing feedback to senior engineers.
Developed a Python function to extract motor operation data and compare motor efficiency, replacing the manual comparison process and improving team efficiency.
Outcome: Successfully completed the assembly, engineering drawings, and data maintenance for the motor sections of three NIO vehicle models. Optimized team workflow using Python, resulting in increased efficiency.
Publications
MambaCPU: Enhanced Correlation Mining with State Space Models for CPU Performance Prediction
ICASSP 2025
Accepted
Abstract
CPU performance prediction, which involves forecasting the performance scores of a CPU based on its hardware characteristics during the operation process, is a critical technology for computational system design and resource management. However, this research field currently faces two significant challenges. First, collecting real-world data is challenging due to the wide variety of CPU products on the market and the highly specialized nature of relevant hardware characteristics. Second, existing methods based on hardware simulation models or machine learning exhibit notable shortcomings, such as lengthy simulation test cycles, low prediction accuracy, and the ignoration of characteristic correlations…
Towards CPU Performance Prediction: New Challenge Benchmark Dataset and Novel Approach
Expert Systems with Applications
Under Review
Abstract
CPU performance prediction, which involves forecasting the performance scores of a CPU based on its hardware characteristics during the operation process, is a critical technology for computational system design and resource management. However, this research field currently faces two significant challenges. First, collecting real-world data is challenging due to the wide variety of CPU products on the market and the highly specialized nature of relevant hardware characteristics. Second, existing methods based on hardware simulation models or machine learning exhibit notable shortcomings, such as lengthy simulation test cycles, low prediction accuracy, and the ignoration of characteristic correlations…
Awards
National First Prize (1st out of 125+ team in China) 2018
National Silver Award (2nd out of 200+ team, prize money ￥50,000)
The National First Prize Scholarship (Top 5%)
The National Third Prize Scholarship (Top 15%)
The National First Prize Scholarship (Top 5%)
Bosch China IoT Hackathon | 2018
National University Automotive Big Data I&E Competition | 2018
HFUT | 2015
HFUT | 2014
HFUT | 2013
Skills
Programming: Python (Proficient), C/C++ (Proficient), MySQL (Experienced)
Framework & Tools: Py-Torch (Proficient), TensorFlow (Proficient), Docker, Hadoop
Research Related: Comprehensive academic writing, Critical thinking, Oral presentation skills.
Language: Chinese (Native), Germany (TestDaF:16), English (IELTS:7.0)