Efficient AI System Design With Cross-Layer Approximate Computing

Venkataramani, Swagath; Sun, Xiao; Wang, Naigang; Chen, Chia-Yu; Choi, Jungwook; Kang, Mingu; Agarwal, Ankur; Oh, Jinwook; Jain, Shubham; Babinsky, Tina; Cao, Nianwen; Fox, Thomas W.; Fleischer, Bruce; Gristede, George; Guillorn, Michael A.; Haynie, Howard; Inoue, Hiroshi; Ishizaki, Kazuaki; Klaiber, Michael J.; Lo, Shih-Hsien; Maier, Gary; Mueller, Silvia Melitta; Scheuermann, M.; Ogawa, Eri; Schaal, Marcel; Serrano, Mauricio J.; Silberman, J. A.; Vezyrtzis, Christos; Wang, Wei; Yee, Fanchieh; Zhang, Jintao; Ziegler, Matthew M.; Zhou, Ching; O’Hara, Maureen; Lu, Pong‐Fei; Curran, Brian; Shukla, Sunil; Srinivasan, Vijayalakshmi; Chang, Leland; Gopalakrishnan, Kailash

doi:10.1109/jproc.2020.3029453

Cited by 45 publications

(29 citation statements)

References 65 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

Section: Part 3: Coupling Of Smart Design and Smart Systemsmentioning

confidence: 99%

“…Therefore, they can be seen both as a constraint on the intellect development process and as a facilitator of tailoring problem solvers. The advantage of using AI-platforms is that application systems can be developed faster and cheaper than when they are built standalone (Venkataramani et al, 2020). AI-frameworks are the conceptual constructs of the cognitive parts of smart systems including logical, functional, architectural, computational, and interaction aspects.…”

mentioning

confidence: 99%

“…Smith (2019) discussed the paradox of being flooded by information about the allmightiness of artificial intelligence as well as the obvious dreads and lasting uncertainties. While positive experience is seen as a success factor of adopting AI in design, Chen et al (2020) reported on the lack of knowledge about the success factors of AI adoption in the telecommunication industry.…”

mentioning

confidence: 99%

See 2 more Smart Citations

Connectors of smart design and smart systems

Horváth

2021

AIEDAM

View full text Add to dashboard Cite

Though they can be traced back to different roots, both smart design and smart systems have to do with the recent developments of artificial intelligence. There are two major questions related to them: (i) What way are smart design and smart systems enabled by artificial narrow, general, or super intelligence? and (ii) How can smart design be used in the realization of smart systems? and How can smart systems contribute to smart designing? A difficulty is that there are no exact definitions for these novel concepts in the literature. The endeavor to analyze the current situation and to answer the above questions stimulated an exploratory research whose first findings are summarized in this paper. Its first part elaborates on a plausible interpretation of the concept of smartness and provides an overview of the characteristics of smart design as a creative problem solving methodology supported by artificial intelligence. The second part exposes the paradigmatic features and system engineering issues of smart systems, which are equipped with application-specific synthetic system knowledge and reasoning mechanisms. The third part presents and elaborates on a conceptual model of AI-based couplings of smart design and smart systems. The couplings may manifest in various concrete forms in real life that are referred to as “connectors” in this paper. The principal types of connectors are exemplified and discussed. It has been found that smart design tends to manifest as a methodology of blue-printing smart systems and that smart systems will be intellectualized the enablers of implementation of smart design. Understanding the affordances of and creating proper connectors between smart design and smart systems need further explorative research.

show abstract

Section: Part 3: Coupling Of Smart Design and Smart Systemsmentioning

confidence: 99%

mentioning

confidence: 99%

mentioning

confidence: 99%

See 1 more Smart Citation

Connectors of smart design and smart systems

Horváth

2021

AIEDAM

View full text Add to dashboard Cite

show abstract

“…Approximate implementations are introduced to CNNs at the level of data type selection, quantization, microarchitecture (e.g., pruning), arithmetic circuits, and memory subsystem. The most significant gains are obtained when adopting a cross-layer approximation approach, which involves software, architecture, and hardware at the same time, breaking thus conventional methods focused on optimizing each layer of abstraction independently [65]. A recent survey of methods for evolutionary design of neural network architectures was presented in [66].…”

Section: Evolutionary Approximation In Cnnsmentioning

confidence: 99%

Evolutionary Algorithms in Approximate Computing: A Survey

Sekanina

2021

JICS

View full text Add to dashboard Cite

In recent years, many design automation methods have been developed to routinely create approximate implementations of circuits and programs that show excellent trade-offs between the quality of output and required resources. This paper deals with evolutionary approximation as one of the popular approximation methods. The paper provides the first survey of evolutionary algorithm (EA)-based approaches applied in the context of approximate computing. The survey reveals that EAs are primarily applied as multi-objective optimizers. We propose to divide these approaches into two main classes: (i) parameter optimization in which the EA optimizes a vector of system parameters, and (ii) synthesis and optimization in which EA is responsible for determining the architecture and parameters of the resulting system. The evolutionary approximation has been applied at all levels of design abstraction and in many different applications. The neural architecture search enabling the automated hardware-aware design of approximate deep neural networks was identified as a newly emerging topic in this area.

show abstract

“…Even though the hardware accelerators might be a solution towards addressing the computing limitations of embedded devices, integrating thousands of MAC units in order to keep up with the computational demands, results in increased energy consumption [2]. Interestingly, previous research works [5]- [10] have shown that a great amount of these computations can tolerate at least some degree of approximation, thus reducing energy consumption and without sacrificing the NN inference accuracy. Thus, exploiting the principle of approximate computing, we can trade-off the system's energy efficiency with respect to the NN accuracy.…”

Section: Introductionmentioning

confidence: 99%

Positive/Negative Approximate Multipliers for DNN Accelerators

Spantidi,

Zervakis,

Anagnostopoulos

et al. 2021

Preprint

View full text Add to dashboard Cite

Recent Deep Neural Networks (DNNs) managed to deliver superhuman accuracy levels on many AI tasks. Several applications rely more and more on DNNs to deliver sophisticated services and DNN accelerators are becoming integral components of modern systems-on-chips. DNNs perform millions of arithmetic operations per inference and DNN accelerators integrate thousands of multiply-accumulate units leading to increased energy requirements. Approximate computing principles are employed to significantly lower the energy consumption of DNN accelerators at the cost of some accuracy loss. Nevertheless, recent research demonstrated that complex DNNs are increasingly sensitive to approximation. Hence, the obtained energy savings are often limited when targeting tight accuracy constraints. In this work, we present a dynamically configurable approximate multiplier that supports three operation modes, i.e., exact, positive error, and negative error. In addition, we propose a filter-oriented approximation method to map the weights to the appropriate modes of the approximate multiplier. Our mapping algorithm balances the positive with the negative errors due to the approximate multiplications, aiming at maximizing the energy reduction while minimizing the overall convolution error. We evaluate our approach on multiple DNNs and datasets against state-of-the-art approaches, where our method achieves 18.33% energy gains on average across 7 NNs on 4 different datasets for a maximum accuracy drop of only 1%.

show abstract

Efficient AI System Design With Cross-Layer Approximate Computing

Cited by 45 publications

References 65 publications

Connectors of smart design and smart systems

Connectors of smart design and smart systems

Evolutionary Algorithms in Approximate Computing: A Survey

Positive/Negative Approximate Multipliers for DNN Accelerators

Contact Info

Product

Resources

About