Domain-Specific Architectures: Research Problems and Promising Approaches

Abstract

Process technology-driven performance and energy efficiency improvements have slowed down as we approach physical design limits. General-purpose manycore architectures attempt to circumvent this challenge, but they have a significant performance and energy-efficient gap compared to special-purpose solutions. Domain-specific architectures (DSAs), an instance of heterogeneous architectures, efficiently combine general-purpose cores and specialized hardware accelerators to boost energy efficiency and provide programming flexibility. Indeed, the hardware, software, and systems aspects in DSAs are highly tailored to maximize the energy efficiency of applications in a target domain. As DSAs and their conceptualization advance rapidly, there is a strong need to understand the research problems that need immediate attention. This paper discusses the primary research directions in the design and runtime management of DSAs. Then, it surveys some promising approaches and highlights the outstanding research needs.

Emerging opportunities and challenges for the future of reservoir computing

“…Research of how to design RC architectures, how to train them and why they work have, over the past two decades following the pioneering works of Jaeger and Maass, led to much evolved view of the capabilities as well as limitations of the RC framework for learning. On the one hand, simulation and numerical research has produced many new network architectures improving the performance of RC beyond purely random connections; future works can either adopt a one-fits-all approach to investigate very large random RCs or perhaps more likely to follow the concept of domain-specific architecture (DSA) 52 to explore structured classes of RCs that achieve optimal performance for particular types of applications, with Bayesian optimization 26 , 27 and NAS as powerful tools of investigation 53 . On the other hand, for a long time only few theoretical guidelines based on ESP were available for practical design of RCs; more recently several important theoretical discoveries were made establishing universal approximation theorems of RC - those results, although not yet directly useful for constructing optimal RCs, may nevertheless boost confidence and stimulate explicitly ideas of designing and even optimizing RCs for learning.…”

M-LAB: scheduling space exploration of multitasks on tiled deep learning accelerators

“…Constrained by the stagnation of Moore's Law [1] , processor capabilities can no longer keep pace with escalating computing demands. Consequently, the focus of research objectives has shifted from obtaining additional computational resources to improving their utilization, culminating in the creation of domain-specific architectures (DSA) [2] .…”

Theoretical Validation and Hardware Implementation of Dynamic Adaptive Scheduling for Heterogeneous Systems on Chip

“…If the scheduler takes a significantly longer amount of time to make a decision, it can undermine the benefits of hardware acceleration. For instance, the Linux Completely Fair Scheduler (CFS) takes 1.2 µs to make a scheduling decision when running on an Arm Cortex-A53 core [8][9][10][11]. This overhead is clearly unacceptable when there are many tasks with orders of magnitude faster execution times.…”