| Title |
Design and Comparison of Radix-4 and Factored Radix-8Modified Booth Multiplier for Neural Processing |
| Authors |
홍태양(Taeyang Hong) ; 정재용(Jaeyong Chung) |
| DOI |
https://doi.org/10.5573/ieie.2020.57.12.81 |
| Keywords |
Deep learning; Machine learning; Modified booth multipliers; Factoring; Pipelining |
| Abstract |
There are many studies on a machine learning accelerator that can process high speed matrix multiplication operations. In this paper, We compare and analyze the recently proposed Factored Radix-8 booth multiplier and conventional Radix-4 booth multiplier focusing on the logic level. Through experiments, we found that the degree and the main cause of performance improvement may vary depending on the different word-lengths. In the booth recording algorithm of Factored Radix-8 multiplier, the most significant bit (MSB) is required to be duplicated at all most word-lengths than Radix-4 multiplier, and has the possibility to consist a multiplier with the same tree level at a specific word-length. Nevertheless, The proposed Factored Radix-8 booth multiplier improves area, delay, ADP up to 10%, 14%, and 23%, respectively, for different word-lengths when compared with the conventional Radix-4 booth multiplier. This shows that the effect at the multiplier level as well as the factoring method can be the main cause of performance improvement. |