Analysing emerging memory technologies for big data and signal processing applications

Abstract

In this paper, we investigate and compare different emerging memory technologies as on-chip cache for big data and signal processing applications. Static Random Access Memory (SRAM) has been widely used as level 1 and last level caches for multicore processors. Server chips integrate Dynamic Random Access Memory (DRAM) as an additional cache for better server-level applications that process more data. Both SRAM and DRAM have advantages and disadvantages. Therefore new types of RAMs are proposed and prototyped. For big data and signal processing applications nowadays, enormous amount of data are processed, usually with time limitations. We analyse novel RAMs, including Phase-change RAM (PRAM), Magnetoresistive RAM (MRAM), Ferroelectric RAM (FRAM) and Resistive RAM (RRAM). The conventional and new memories are analysed in terms of size, area, access latency and power consumption. We present benchmark results using a full system simulator. Workloads are selected from several big data, server, signal processing and video processing applications. Experiments show that, in consideration of these applications, it is crucial to replace SRAM and DRAM caches with MRAM and RRAM.