National Technical University of Athens (NTUA)
From circuits to SoC processors: arithmetic approximation techniques and embedded computing methodologies for DSP acceleration
Abstract
dc:descriptionThe recent end of Dennard's Scaling and the declining Moore's Law have signified a new era for the computing systems. Power efficiency has now become a critical factor for both cloud and edge computing. Concurrently, the rapid growth of compute-intensive applications from the Digital Signal Processing (DSP) and Artificial Intelligence (AI) domains challenges the resources of computing systems. As a result, the computing industry is forced to find alternative design approaches and computing platforms to sustain increased power efficiency, while providing sufficient performance. Among the examined solutions, Approximate Computing, Hardware Acceleration, and Heterogeneous Computing have gained great momentum. Approximate Computing is a novel design paradigm that exploits the inherent error resilience of DSP/AI applications to deliver gains in power, area, and/or performance by reducing the quality of the results. Hardware Acceleration refers to the execution of demanding computational tasks on specialized hardware, such as the Application-Specific Integrated Circuits (ASICs) and the Field-Programmable Gate Arrays (FPGAs), rather than general-purpose processors. Finally, Heterogeneous Computing refers to versatile processing architectures, such as the Vision Processing Units (VPUs), which integrate more than one type of processor and various memory technologies. In this Dissertation, we introduce design solutions and methodologies, built on top of the preceding computing paradigms, for the development of energy-efficient DSP and AI accelerators. In particular, we adopt the promising paradigm of Approximate Computing and apply new approximation techniques in the design of arithmetic circuits. Based on our methodology, these arithmetic approximation techniques are then combined with hardware design techniques to implement approximate ASIC- and FPGA-based DSP and AI accelerators. Moreover, we propose methodologies for the efficient mapping of DSP/AI kernels on distinctive embedded devices, such as the new space-grade FPGAs and the heterogeneous VPUs. On the one hand, we cope with the decreased flexibility of the space-grade technology and the technical challenges that arise in new FPGA tools and devices. On the other hand, we unlock the full potential of heterogeneity by surpassing the increased hardware complexity and exploiting all the diverse processors and memories. In more detail, the proposed arithmetic approximation techniques involve bit-level optimizations, inexact operand encodings, and skipping of computations, while they are applied in both fixed- and floating-point arithmetic. To increase the design space and extract the most efficient solutions, we also conduct an extensive exploration on combinations among the approximation techniques. Moreover, we propose a low-overhead scheme for seamlessly adjusting the approximation degree of our circuits at runtime. In comparison with state-of-the-art designs, the proposed arithmetic circuits feature a very large approximation space, i.e., a wide range of approximation configurations, which enable to maximize the resource gains for a given error constraint. Our techniques induce a mean relative error of up to ~2%, i.e., typical error values for approximate circuits. The most prominent approximate circuits of the Dissertation form a high-resolution Pareto front in a comparative evaluation involving state-of-the-art designs of the literature, and they deliver up to 63% better energy consumption. Finally, our runtime-configurable circuits exhibit a small area overhead of ~3% compared to the accurate design, and they provide ~1.5× less energy gains than their respective design-time counterparts with fixed approximation. Nevertheless, they can dynamically change the approximation degree, namely, the accuracy of the calculations, while they still attain remarkable energy gains versus the accurate circuit and state-of-the-art approximate circuits. At the accelerator level, we develop a plethora of approximate kernels for 1D/2D signal processing and Convolutional Neural Networks (CNNs). The experimental results show that we achieve small relative errors for classic DSP calculations and 0%-5% accuracy loss in CNNs for various arithmetic formats, while providing up to 70% area and energy savings. Regarding the DSP acceleration on new space-grade FPGAs, we apply our methodology to efficiently map computer vision algorithms onto the radiation-hardened NanoXplore's FPGAs. In the end, we achieve balanced resource utilization, which is comparable to that of well-established FPGA vendors. Furthermore, the throughput is sufficient (e.g., up to 10 FPS for feature detection on MPixel images), considering the performance requirements of vision-based space applications. In terms of Heterogeneous Computing, we accelerate custom DSP kernels, a sophisticated computer vision pipeline, and a demanding CNN with ResNet-50 backbone on Intel's Myriad VPUs. The proposed methodology and embedded design techniques provide speedups up to 20× for classic DSP on Myriad 2, while the power lies around 1W. The CNN is accelerated on Myriad X with 2W, achieving ~8.5× and ~1.7× better performance-per-Watt than the ARM CPU and the Jetson Nano GPU, respectively.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2022
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Leon, Vasileios
- Λέων, Βασίλειος
Subjects
dc:subject × 28- Προσεγγιστικός υπολογισμός
- Τεχνικές προσέγγισης
- Αριθμητική υπολογιστών
- Αριθμητικά κυκλώματα
- Σχεδίαση υλικού
- Επιταχυντές υλικού
- Ετερογενής υπολογισμός
- Ενσωματωμένα συστήματα
- Τεχνολογία διαστήματος
- Ψηφιακή επεξεργασία σήματος
- Υπολογιστική όραση
- Approximate computing
- Approximation techniques
- Computer arithmetic
- Arithmetic circuits
- Hardware design
- Hardware accelerators
- Heterogeneous computing
- Embedded systems
- Space-grade
- Digital signal processing
- Computer vision
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Υπολογιστές, Υλικό (hardware) και Αρχιτεκτονική
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Computer science, Hardware and Architecture
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/52666