A member of Performance team, responsible for optimizing Microsoft/M365 resources usage.

Department: Alibaba Group-DAMO-Autonomous Driving Lab-Computing PlatformResponsible for developing & optimizing GPU autonomous driving platform.

a member of TensorRT team.

Meh

1) 2015-: Optimization Engineer at Tencent(Shenzhen) a) A distributed logistic regression i. Combined MPI/OpenMP/SIMD with a highly tuned RLE encode/decode, the whole system contains 14nodes X 16 CPUs. It is capable of handling up to 1T data. Compared to the legacy one, it's 40x times faster and uses only quarter of memory resources. b) A highly optimized FTRL logistic regression i. Since FTRL is a strictly sequential, low computing-intensity algorithm, this system adapted a totally different approach to boost its efficiency. With a Out-Of-Order software pipeline, it can execute downloading, decompressing, parsing data and other tasks in parallel, after all of those steps, the computing and updating of weight would be executed strictly sequential. c) A couple of discontinued Machine Learning projects i. Including a CNN for image classification, an AI for Texas HoldEm poker, an MCST+DCNN robot for Go, etc.

1) 2014 – 2015 Compiler Engineer at Intela) develop and maintain several components of ART(a head-of-ti me Java compiler for Android).b) Bug fixing and enhancing for Render Scripts, which is an Android tool set similar to OpenCL, but renowned for its poor design.2) 2011 – 2013 Tools Engineer at Intela) An Android evaluation tool based on CUDAi. By deploying 3 TiTan/CUDA cards, the system can handle 0.5GB image and log data per second.ii. The overall processing time reduced from 20 Minutes to 4 seconds.b) An Android system image flashing and setting tooli. Develop and maintain a easy-to-use internal/external tool for flashing android phones. It’s very popular in Intel branches globally, and the total number of user adds up to more than 6000+.ii. A couple of other tools for Android, but them are not as popular as the one mentioned above.