Profile

AI Accelerators · Reinforcement Learning · FPGA/NPU

Wei Xiangqing

I am a doctoral researcher focusing on hardware–software co-design for deep learning accelerators. My current work builds closed-loop reinforcement-learning-driven scheduling frameworks for NPU and FPGA-based neural network accelerators.

Wei Xiangqing profile photo

Research Focus

RL-driven Accelerator Scheduling

I study reinforcement-learning-based scheduling methods for neural network accelerators, focusing on tile dispatch, PE allocation, latency reduction, energy efficiency, and utilization improvement.

Tile-level DAG Modeling

I model CNN workloads as tile-level directed acyclic graphs to capture computation dependencies, weight reuse, memory conflicts, and hardware execution constraints.

NPU and FPGA Acceleration

My work involves NVDLA-style and Gemmini-style accelerator modeling, MAC/PE-array analysis, RTL-level instrumentation, and FPGA-based validation.

Memory-aware Optimization

I am interested in RL-driven memory optimization for sparse neural models, including prefetching, bank allocation, data reuse, and irregular memory-access scheduling.

News

Coming soon.
I will update this section with research progress, academic activities, and website updates.

Academic Background

Research Assistant

Chiba University, Japan
Research area: AI accelerators, NPU scheduling, reinforcement learning, FPGA-based accelerator design.

Master’s Degree in Integrated Circuit Engineering

Guizhou University, China
Research area: RISC-V processor design, integer execution units, Booth multiplication, SRT division, and digital circuit design.

Research Keywords

AI Accelerators NPU FPGA Reinforcement Learning Tile-level Scheduling DAG Modeling Sparse Neural Networks Hardware-Software Co-design Computer Architecture RTL Design

Current research direction: I am building structure-aware and memory-aware scheduling frameworks for deep learning accelerators, aiming to bridge neural network workload modeling, reinforcement learning, and hardware-level execution feedback.