Design Methodology And Electronic Design Automation Techniques For Heterogeneous 3D Machine Learning Accelerators
Loading...
Author(s)
Murali, Gauthaman
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
The primary goal of this thesis is to present comprehensive design methodologies and
electronic design automation (EDA) techniques aimed at enhancing the efficiency of 3D
machine learning accelerator designs. This research extensively investigates the challenges
associated with scaling near-memory and in-memory compute accelerators in the X/Y dimensions
and proposes optimized scaling solutions by leveraging heterogeneous 3D integration.
While numerous 3D physical design EDA methodologies exist, they often involve a
multitude of parameters that require careful tuning based on the specific design in question.
To address this, the thesis examines the impact of these parameters on various design types
using cutting-edge machine learning-based parameter autotuning tools. It subsequently
recommends the most effective parameter tuning techniques depending on the particular
3D EDA approach being employed.
Using the EDA tools optimized through machine learning, the thesis delves into the
advantages of heterogeneous 3D implementation for near-memory and in-memory compute
ML accelerators at various technology nodes. Heterogeneous 3D integration delivers
substantial improvements in terms of Power-Performance-Area (PPA), throughput, energy
efficiency, and area efficiency when compared to their 2D counterparts.
Finally, the thesis introduces a straightforward yet highly efficient machine learning
framework for exploring the design space of heterogeneous 3D accelerators. This framework
assists designers in selecting the optimal accelerator architecture tailored to the requirements
of their target applications.
Sponsor
Date
2023-11-09
Extent
Resource Type
Text
Resource Subtype
Dissertation