UM Professor Ka-Fai Un Gives a Talk on Energy-efficient Reconfigurable CNN Accelerator for Object Recognition Applications


Despite FPGA’s advantages of high configurability and robustness, its energy efficiency is far from satisfactory, which hurdles its wide adoption. Prof. Ka-Fai Un, Assistant Professor, State Key Laboratory of Analog and Mixed-Signal VLSI and FST-ECE, University of Macau (UM) has probed into the issue and shared his research result with UMEC today.

The objective of Prof. Un’s research is to develop a FPGA-based CNN accelerator with high throughput and energy efficiency. By adopting the kernel partition technique, the accelerator can balance the ability of parallel computing with reduced memory access to the input feature maps and the CNN kernels. Besides, kernel stack technique is used to substantially reduce the buffers for storing partial sums. The experimental results showed that it can achieve 1.6X energy efficiency than other FPGA accelerators, while preserving high throughput and high computation, which make it an ideal solution for real-time object recognition applications.

UMEC’s colleagues were excited about the research result. Multiple questions had been raised to further understand the architecture and applications of the CNN accelerator. Much work could be done between Prof. Un and UMEC to facilitate its advancement and applications, in order to drive extensive commercialisation.


Prof. Ka-Fai Un shared the FPGA-based energy-efficient reconfigurable CNN accelerator for object recognition applications