A Data-Centric Accelerator Design Based on Processing in Memory

Abstract

The data-centric computing paradigm has recently garnered a great deal of attention from the research community as a method for overcoming the performance limits of traditional computing systems, including the memory wall crisis. One promising approach to mitigating this issue in future computer systems is processing in memory (PIM). PIM facilitates the stacking of processing logic and memory dies in a single package and minimizes data movement by placing the computation close to where the data reside. As this approach, however, requires compatibility with existing computer architectures and operating systems, it has not been widely adopted. To meet the need for compatibility, this paper proposes a hardware architecture of PIM and verifies the functions of the proposed architecture by an embedded system based on the PIM platform, which employs a commercialized application processor (AP) and a standard memory protocol. We also propose a PIM-based data-centric accelerator for image processing. Experiments involve the development of AP- and PIM-based application programs for processing a median filter that uses a 24-bit color image with a 512 × 512 resolution test image. Using the same test image, we measure the median filter processing times and compare the processing times of the AP and the proposed PIM. Results of the experiments show that the processing time of PIM is about 84% faster than that of the AP.