부산대학교 도서관

To reduce the impact of the memory-access constraint in k-Nearest Neighbors (kNN) problems, in this paper we implement one kNN kernel through high-level synthesis (HLS) on FPGA by employing two data access reduction methods: low-precision data representation and principal component analysis based filtering (PCAF). The kernel is called MPCAF-kNN (Memory-efficient PCAF kNN), which has been highly optimized to fully exploit the characteristics of FPGA. It is adaptive to all key parameters. We evaluate MPCAF-kNN by comparing it with a state-of-the-art kNN implementation on a high-end CPU server. Our results show that MPCAF-kNN achieves up to a performance equivalent to that of a 56-thread of CPU server while greatly reducing external memory-accesses.