You are here

PhD Defense: Hamed TabkhiVayghan

11
Aug

442 Dana

August 11, 2014 3:30 pm to 5:30 pm
August 11, 2014 3:30 pm to 5:30 pm

Hamed TabkhiVayghan presents his Ph.D Defense

 

Date:  August 11, 2014

Time: 3:30PM-5:30PM.

Location: 442 Dana

 

Advisor: Professor Gunar Schirner

Committee members: Professor David Kaeli and Professor Miriam Leeser

 

Title: High-Performance Power-Efficient Solutions for Pervasive Embedded Vision Computing

 

Embedded vision is recognized as a top tier, rapidly growing market with a host of challenges and conflicting requirements. Embedded vision platforms often set to run complex algorithms with immense computation/communication (50 – 100 GOPs/8 – 10 GBPs) at low power (few Watts).

To support the vision market requirements and mitigate the design and development costs, heterogeneous Multiprocessor System-on-Chips (MPSoCs) have emerged as the main architecture solutions. However, current vision MPSoCs either shy away from supporting advanced vision algorithms or perform very limited resolutions due to not tackling the immense communication/computation demands of vision algorithms.

This dissertation identifies three major challenges hindering the embedded realizations of advanced vision algorithms: (a) design gap, (b) architecture gap and (c) abstraction gap. The design gap (a) is a gap between the vision algorithm requirements and SoC capabilities. The advanced vision algorithms are focusing on complex scene modeling requiring massive communication demands which are often far beyond the SoC design capabilities. The architecture gap (b) is a gap between flexibility (as Instruction-level-Processors) and efficiency (as Custom-HWs). Since, designing custom-HWs for every vision kernel is cost prohibitive, large portions of vision applications still run inefficiently on ILPs. The abstraction gap (c) is a gap between market requirements and system specification model. System architects often start from system specification model and rely on their evolving knowledge to architect vision platforms. A definition of abstraction levels and automation tools are required to guide system architects starting from market requirement down to system specification.

This dissertation proposes three main contributions to address the identified challenges:

(a) traffic separation in adaptive vision algorithms,

(b) Function-Level-Processor (FLP) and

(c) Conceptual-Abstraction-Levels (CALs).

The traffic separation (a) identifies two class of data access patterns: streaming and algorithm-intrinsic. The traffic separation enables application-specific management of algorithm-intrinsic data (compression, prioritization), and simplifies pipelining of stream processing nodes – reducing the immense communication demand and removes a tremendous hurdle for embedded realization of adaptive vision algorithms. The power / performance benefits of proposed architecture-template is demonstrated by constructing a complete object tracking vision flow on Zynq-based architecture executing 50GOPs at 1.5Watts of on-chip power. FLP (b) offers a new architecture alternative in the flexibility / efficiency trade-off by providing efficiency and performance similar to custom-HWs and yet flexibility to execute different applications. The FLP efficiency stems from coarsening architecture programmability from instructions (as in ILPs) to function blocks. An early version of the FLP (ADIs PVP) processes up to 22.4 GOPs consuming 314 mWatts – 14x-18x less than comparable ILP-based solutions. CALs (c) bridge the gap between market requirements and system specification model. CALs identify a sequence of critical areas for early architecture exploration and resolve interdependent challenges and dependencies through iteration. CALs help system architects to identify the potential application taking benefits of traffic separation or application blocks for function-level processing at early stages of design.

Overall, the aim of this dissertation is to tackle associated architecting / design complexities of embedded vision MPSoCs. In a holistic view, the dissertation contributions in principle are applicable to other challenging market domains featuring stream data processing such as radars, wireless base-band processing and cyber-physical systems.