• As a core developer of the NPU compiler, I was deeply involved in and designed the first-generation NN compiler (codenamed MCM). I implemented the automatic compilation, optimization, and blob generation of neural network models under the OpenVINO IR. In 2021, I migrated it into the MLIR ecosystem and led the development of the first automatic spatial tiling strategy allocation algorithm, significantly improving the performance of CI models. This algorithm has been continuously iterated and used ever since. I was the first to introduce the NN cost model into the compiler, establishing a cost-based tiling strategy evaluation mechanism. This enabled the quantitative selection of the optimal strategy, which has been continuously expanded and improved to this day, becoming one of the key advanced features of the compiler.
• I have been involved in the overall architecture design and maintenance of the NPU compiler, as well as the implementation of key features for each generation. As a core contributor and maintainer of the codebase, I have been responsible for the long-term review and merging of PRs (pull requests).
• I have been responsible for performance tuning of customer models. Taking Microsoft as an example, I helped Microsoft's PS/MEP/Copilot+ models achieve the best performance and experience on Intel NPUs, including optimizing model compilation time, fixing model output accuracy, and analyzing and optimizing model performance bottlenecks. With the joint efforts of the global NPU team, we defeated Qualcomm to become Microsoft's primary partner!

The NPU products involved in the above work include Keembay, Meteor Lake, Lunar Lake, and Panther Lake. Keembay is the second-generation standalone product, while NPUs from the third generation onwards are integrated into Intel Core Ultra processors, similar to how iNPU is akin to integrated graphics iGPU.