{"title":"A 28.5mW 2.8GFLOPS floating-point multifunction unit for handheld 3D graphics processors","authors":"Byeong-Gyu Nam, H. Yoo","doi":"10.1109/ASSCC.2007.4425709","DOIUrl":null,"url":null,"abstract":"A low-power, high-performance 4-way 32-bit floating-point multifunction unit is developed for handheld 3D graphics processors. It uses logarithmic arithmetic to unify matrix, vector, and elementary functions into a single arithmetic unit. The optimal designs of logarithmic and antilogarithmic converters are presented. An adaptive number conversion scheme is proposed and it reduces total area by 15%. With this scheme, the matrix-vector multiplication (MAT), cross-product, lerp, and logarithm (logx y with 2 variables) are newly unified with the other operations. The unit achieves 2-cycle throughput for the MAT and single-cycle throughput for all other operations. It takes 451 K transistors and achieves 2.8 GFLOPS at 200 MHz with 28.5 mW power consumption.","PeriodicalId":186095,"journal":{"name":"2007 IEEE Asian Solid-State Circuits Conference","volume":"8 1","pages":"0"},"PeriodicalIF":0.0000,"publicationDate":"2007-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"5","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"2007 IEEE Asian Solid-State Circuits Conference","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1109/ASSCC.2007.4425709","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 5
Abstract
A low-power, high-performance 4-way 32-bit floating-point multifunction unit is developed for handheld 3D graphics processors. It uses logarithmic arithmetic to unify matrix, vector, and elementary functions into a single arithmetic unit. The optimal designs of logarithmic and antilogarithmic converters are presented. An adaptive number conversion scheme is proposed and it reduces total area by 15%. With this scheme, the matrix-vector multiplication (MAT), cross-product, lerp, and logarithm (logx y with 2 variables) are newly unified with the other operations. The unit achieves 2-cycle throughput for the MAT and single-cycle throughput for all other operations. It takes 451 K transistors and achieves 2.8 GFLOPS at 200 MHz with 28.5 mW power consumption.