{"title":"An 8Kbyte intelligent cache memory","authors":"T. Watanabe","doi":"10.1109/ISSCC.1987.1157109","DOIUrl":null,"url":null,"abstract":"WITH THE ADVENT OF FIKE-LINE fabrication technology and incorporation of pipelining architecture, recent single-chip microprocessors have maximum speed as high as several MIPS. On the other hand, the speed of the system bus is relatively low because of slower main memories. To fill this speed gap, the use of cache memories in a microprocessor system has become commonplace. Some modern microprocessors have on-chip memories, and one related paper’ introduced a dedicated chip for the cache memory and memory management unit. A general-purpose intelligent cache memory with 8Kbyte data memories and support functions, will be reported. Four design goals were: compatibility with most high-performance 16b and 32b microprocessor^^'^'^, no-wait cycles at MPU clock rates of 16MHz or higher, cache-miss ratio of less than 57’0, and expandability to a multi-processor system. The first two requirements require the cache memory to operate at an access time of less than 70ns from address strobe signal to ready signal, and at an access time of less than 85ns from address strobe signal to data valid. Tradeoffs between including a larger data memory and integrating more control functions in the cache chip were examined by computer simulation to obtain a solution to the cache-miss ratio. Simulation showed that a data memory of more than 8Kbytes should be introduced on one chip, even if sophisticated housekeeping techniques could be adopted. The best approach was revealed to be the combination of a 4-way set associative placement algorithm, 16bytes block size, L R U replacement algorithm, and use of pre-fetch.","PeriodicalId":102932,"journal":{"name":"1987 IEEE International Solid-State Circuits Conference. Digest of Technical Papers","volume":"8 1","pages":"0"},"PeriodicalIF":0.0000,"publicationDate":"1900-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"14","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"1987 IEEE International Solid-State Circuits Conference. Digest of Technical Papers","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1109/ISSCC.1987.1157109","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 14
Abstract
WITH THE ADVENT OF FIKE-LINE fabrication technology and incorporation of pipelining architecture, recent single-chip microprocessors have maximum speed as high as several MIPS. On the other hand, the speed of the system bus is relatively low because of slower main memories. To fill this speed gap, the use of cache memories in a microprocessor system has become commonplace. Some modern microprocessors have on-chip memories, and one related paper’ introduced a dedicated chip for the cache memory and memory management unit. A general-purpose intelligent cache memory with 8Kbyte data memories and support functions, will be reported. Four design goals were: compatibility with most high-performance 16b and 32b microprocessor^^'^'^, no-wait cycles at MPU clock rates of 16MHz or higher, cache-miss ratio of less than 57’0, and expandability to a multi-processor system. The first two requirements require the cache memory to operate at an access time of less than 70ns from address strobe signal to ready signal, and at an access time of less than 85ns from address strobe signal to data valid. Tradeoffs between including a larger data memory and integrating more control functions in the cache chip were examined by computer simulation to obtain a solution to the cache-miss ratio. Simulation showed that a data memory of more than 8Kbytes should be introduced on one chip, even if sophisticated housekeeping techniques could be adopted. The best approach was revealed to be the combination of a 4-way set associative placement algorithm, 16bytes block size, L R U replacement algorithm, and use of pre-fetch.
随着流水线制造技术的出现和流水线架构的结合,最近的单芯片微处理器的最高速度高达几MIPS。另一方面,由于主存储器较慢,系统总线的速度相对较低。为了填补这一速度差距,在微处理器系统中使用高速缓存存储器已经变得司空见惯。一些现代微处理器有片内存储器,一篇相关论文介绍了一种用于缓存存储器和存储器管理单元的专用芯片。将报道具有8Kbyte数据存储和支持功能的通用智能缓存存储器。四个设计目标是:兼容大多数高性能16b和32b微处理器^^'^'^,在16MHz或更高的MPU时钟速率下无等待周期,缓存丢失率小于57 ' 0,以及可扩展到多处理器系统。前两个要求要求缓存存储器从地址频闪信号到就绪信号的访问时间小于70ns,从地址频闪信号到有效数据的访问时间小于85ns。通过计算机仿真研究了在缓存芯片中加入更大的数据存储器和集成更多的控制功能之间的权衡,得到了缓存缺失率的解决方案。模拟表明,即使采用复杂的内务管理技术,一个芯片上也应该引入8kb以上的数据存储器。最好的方法是结合4路集合关联放置算法、16字节块大小、L R U替换算法和使用预取。