Approximate computing techniques for accelerating compute intensive workloads

Kulkarni, Vandana V

Permalink

https://hdl.handle.net/2142/108345 Copy

Description

Title

Approximate computing techniques for accelerating compute intensive workloads

Author(s)

Kulkarni, Vandana V

Issue Date

2020-05-12

Director of Research (if dissertation) or Advisor (if thesis)

Hwu, Wen-Mei W

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

Approximate Computing, Cost Model

Abstract

High Performance Computing involves improving the computational performance of memory and compute intensive workloads in science or engineering. One of the main components to the current success of ML is the ability to perform computations on very large amounts of training data. Similarly in the AR/VR world, high performance computing enables users to interact with large virtual environment and their data at real time. In this thesis, we have focused on Machine Learning Clustering Algorithms and AR/VR SLAM Algorithm. As the complexity of modern computing systems and applications increases, creating highly adaptable programs becomes a very critical and challenging problem. At a high level, there are many algorithmic clustering choices for each application. We have to visualize the data, analyze what the data distribution is and then apply appropriate clustering algorithms based on the data distribution to get good accuracy and performance for a learning problem. KMeans, DBSCAN, etc. are a few choices for clustering algorithms and the ideal choice of the algorithm is input or architecture dependent. An algorithm that is designed for one kind of model will generally fail on a data set that contains a radically different kind of model. We have developed an end-to-end pipeline to choose the most optimal clustering algorithm with the least run time without having to visualize the data. Virtual and Augmented Reality (AR/VR) or more generally Extended Reality (XR) is fast emerging to be a widely used consumer facing technology. Recent advances at the hard- ware and software level are attempting to provide close-to-real virtual reality experience. Extended reality applications incorporate a combination of computational kernels that have varying characteristics. To enable realistic experiences, XR platforms need several orders of magnitude improvements in power and performance. To bridge this gap, we propose to leverage Approximate Computing approaches that trade-off some acceptable accuracy in exchange for significant improvements in energy and performance. Approximate computing approaches have been employed both at the architecture level and the software level particularly in domains that are tolerant to small amounts of error. In this work, we focus on two computational techniques frequently used in the XR domain - SLAM (used in pose prediction) and Object detection (useful in gesture recognition). The focus of our work is to use approximate computing techniques to accelerate the computation of these kernels with graceful degradation in accuracy.

Graduation Semester

2020-05

Type of Resource

Thesis

Permalink

http://hdl.handle.net/2142/108345

Copyright and License Information

Copyright 2020 Vandana Kulkarni

Owning Collections