Files in this item



application/pdfECE499-Sp2019-liu.pdf (1MB)Restricted to U of Illinois
(no description provided)PDF


Title:Pydima: Evaluation on machine learning algorithms on deep in-memory architecture (DIMA)
Author(s):Liu, Siyu
Contributor(s):Shanbhag, Naresh R.
machine learning
always-on inference
process variations
circuit simulation
Abstract:While machine learning delivers state-of-the-art accuracy on many AI tasks, its performance has been limited by the cost of high energy consumption for memory bit-cell-array (BCA) usage. To address this problem, deep in-memory architecture (DIMA) has been designed to enhance both energy efficiency and throughput over conventional digital architectures by reading multiple bits per cycle and by employing mixed-signal processing in the periphery of the BCA. Due to its mixed-signal nature, DIMA is sensitive to transistor-level spatial process variation and data-dependent non-idealities, which circuit simulation tools fail to capture. In order to have an accurate simulation of DIMA, an emulator is developed based on a model which considers the discharge transistor current equation and normal distribution of spatial threshold voltage variation. The result of the emulator is compared with the measurement of a 65 nm CMOS prototype IC running support vector machine (SVM) on face detection dataset from The Center for Biological & Computational Learning (CBCL) at MIT and predicted the inference accuracy with an error range less than 2%.
Issue Date:2019-05
Date Available in IDEALS:2019-06-14

This item appears in the following Collection(s)

Item Statistics