Energy-efficient neural networks using approximate computation reuse

Jiao, Xun; Akhlaghi, Vahideh; Jiang, Yu; Gupta, Rajesh K.

doi:10.23919/date.2018.8342202

Cited by 43 publications

(16 citation statements)

References 17 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Furthermore, custom hardware accelerators are often associated with proposed computational reuse schemes. DNN is known to be error-tolerant; therefore, it is resilient to approximating inputs and weights by reading the binary representation [22], [23]. Such property is leveraged by computational reuse, supporting architectures, and algorithms to increase reuse potential further.…”

Section: A Hardware-based Cnn Acceleratorsmentioning

confidence: 99%

“…Face recognition is a CNN application commonly used in mobile devices that can take advantage of computation reuse due to the high similarity profile of face images. In [23], researchers use similar feature skipping and tile-based hierarchical clustering resulting in a 30% reduction of face recognition CNN computation with only 1% accuracy loss. The proposed ultra-low-power face recognition processors use 1-bit binary weights, which decreases the memory weight footprint and simplifies the MAC processing resulting in very high energy efficiency.…”

Section: A Hardware-based Cnn Acceleratorsmentioning

confidence: 99%

See 1 more Smart Citation

SLID: Exploiting Spatial Locality in Input Data as a Computational Reuse Method for Efficient CNN

et al. 2021

View full text Add to dashboard Cite

Convolutional Neural Networks (CNNs) revolutionized computer vision and reached the state-of-the-art performance for image processing, object recognition, and video classification. Even though CNN inference is notoriously compute-intensive, as convolutions account for > 90% of the total operation tasks, the ability to tradeoff between accuracy, performance, power, and latency to meet target application makes it an open research topic. This paper proposes the Spatial Locality Input Data (SLID) method for computational reuse during the inference stage for a pre-trained network. The method exploits input data spatial locality via skipping partial processing of the multiply-and-accumulate (MAC) operations for adjacent data and equating its value to previously computed ones. SLID improves the throughput of resource-constrained devices (Internet-of-Things, edge devices) and accelerates computations during the inference phase by reducing the number of MAC operations. Such approximate computing schema does not require a similarity quantification step nor any modification for the training stage. The computational data reuse was evaluated on three well-known distinctive CNN structures and data sets with alternating layer selections: LeNet, CIFAR-10, and AlexNet. The computational data reuse method saves up to 34.9%, 49.84%, and 31.5% of MAC operations while reducing the accuracy by 8%, 3.7%, and 5.0% for the three models mentioned earlier, respectively. Besides, the proposed method saves on memory access by eliminating data fetching of skipped inputs. Furthermore, filter size, strides, and padding on the accuracy and savings of operations are analyzed. SLID is the first work to exploit the input spatial locality for savings on CNN convolution operations with minimal accuracy loss and without memory or computational overhead. This makes it a great option to support intelligence at the edge.

show abstract

Section: A Hardware-based Cnn Acceleratorsmentioning

confidence: 99%

Section: A Hardware-based Cnn Acceleratorsmentioning

confidence: 99%

SLID: Exploiting Spatial Locality in Input Data as a Computational Reuse Method for Efficient CNN

et al. 2021

View full text Add to dashboard Cite

show abstract

“…RMAC uses addition for manitssae product computation, instead of multiplying them [17]. In [18], Jiao et al proposed a multiplier which exploits the computation reuse opportunities and enhance them by performing approximate pattern matching.…”

Section: Related Workmentioning

confidence: 99%

AxCEM: Designing Approximate Comparator-Enabled Multipliers

Ghabraei

Rezaalipour

Dehyadegari

et al. 2020

JLPEA

View full text Add to dashboard Cite

Floating-point multipliers have been the key component of nearly all forms of modern computing systems. Most data-intensive applications, such as deep neural networks (DNNs), expend the majority of their resources and energy budget for floating-point multiplication. The error-resilient nature of these applications often suggests employing approximate computing to improve the energy-efficiency, performance, and area of floating-point multipliers. Prior work has shown that employing hardware-oriented approximation for computing the mantissa product may result in significant system energy reduction at the cost of an acceptable computational error. This article examines the design of an approximate comparator used for preforming mantissa products in the floating-point multipliers. First, we illustrate the use of exact comparators for enhancing power, area, and delay of floating-point multipliers. Then, we explore the design space of approximate comparators for designing efficient approximate comparator-enabled multipliers (AxCEM). Our simulation results indicate that the proposed architecture can achieve a 66% reduction in power dissipation, another 66% reduction in die-area, and a 71% decrease in delay. As compared with the state-of-the-art approximate floating-point multipliers, the accuracy loss in DNN applications due to the proposed AxCEM is less than 0.06%.

show abstract

“…Lookup-based approximation avoids computation entirely by reusing previously computed values. Ternary content addressable memories (TCAMs) can be utilized for approximate computational reuse for GPGPU applications [15,16,23,24]. Associative memory is placed adjacent to FPUs and stores input and output values from previously computed operations which are searched against incoming inputs.…”

mentioning

confidence: 99%

“…The authors in [15] proposed a configurable memory which applies VOS to non-volatile associative memory to relax computation and to trade output accuracy for energy savings. Work in [23] designed a novel associative memory based bloom filters. Machine learning algorithms and neural networks have proven to be resilient to some level of reduced precision.…”

mentioning

confidence: 99%

Arga

Peroni

Imani

Nejatollahi

et al. 2019

Proceedings of the 56th Annual Design Automation Conference 2019

View full text Add to dashboard Cite

Many data-driven applications including computer vision, speech recognition, and medical diagnostics show tolerance to error during computation. These applications are often accelerated on GPUs, but high computational costs limit performance and increase energy usage. In this paper, we present ARGA, an approximate computing technique capable of accelerating GPGPU applications. ARGA provides an approximate lookup table to GPGPU cores to avoid recomputing instructions with identical or similar values. We propose multi-table parallel lookup which enables computational reuse to significantly speed-up GPGPU computation by checking incoming instructions in parallel. The inputs of each operation are searched for in a lookup table. Matches resulting in an exact or low error are removed from the floating point pipeline and used directly as output. Matches producing highly inaccurate results are computed on exact hardware to minimize application error. We simulate our design by placing ARGA within each core of an Nvidia Kepler Architecture Titan and an AMD Southern Island 7970. We show our design improves performance throughput by up to 2.7× and improves EDP by 5.3× for 6 GPGPU applications while maintaining less than 5% output error. We also show ARGA accelerates inference of a LeNet NN by 2.1× and improves EDP by 3.7× without significantly impacting classification accuracy. CCS CONCEPTS • Computer systems organization → Multicore architectures; • Computing methodologies → Machine learning approaches.

show abstract

Energy-efficient neural networks using approximate computation reuse

Cited by 43 publications

References 17 publications

SLID: Exploiting Spatial Locality in Input Data as a Computational Reuse Method for Efficient CNN

SLID: Exploiting Spatial Locality in Input Data as a Computational Reuse Method for Efficient CNN

AxCEM: Designing Approximate Comparator-Enabled Multipliers

Arga

Contact Info

Product

Resources

About