Memory-centric reconfigurable accelerator for classification and machine learning applications

Abstract

Big Data refers to the growing challenge of turning massive, often unstructured datasets into meaningful, organized, and actionable data. As datasets grow from petabytes to exabytes and beyond, it becomes increasingly difficult to run advanced analytics, especially Machine Learning (ML) applications, in a reasonable time and on a practical power budget using traditional architectures. Previous work has focused on accelerating analytics readily implemented as SQL queries on data-parallel platforms, generally using off-the-shelf CPUs and General Purpose Graphics Processing Units (GPGPUs) for computation or acceleration. However, these systems are general-purpose and still require a vast amount of data transfer between the storage devices and computing elements, thus limiting the system efficiency. As an alternative, this article presents a reconfigurable memory-centric advanced analytics accelerator that operates at the last level of memory and dramatically reduces energy required for data transfer.We functionally validate the framework using an FPGA-based hardware emulation platform and three representative applications:Naive Bayesian Classification, Convolutional Neural Networks, and k-Means Clustering. Results are compared with implementations on a modern CPU and workstation GPGPU. Finally, the use of in-memory dataset decompression to further reduce data transfer volume is investigated. With these techniques, the system achieves an average energy efficiency improvement of 74×and 212× over GPU and single-threaded CPU, respectively, while dataset compression is shown to improve overall efficiency by an additional 1.8× on average.