Explore chapters and articles related to this topic
Applications of the Approximate Computing on ML Architecture
Published in Sandeep Saini, Kusum Lata, G.R. Sinha, VLSI and Hardware Implementations Using Modern Machine Learning Methods, 2021
Kattekola Naresh, Shubhankar Majumdar
Approximate computation was often used along with NN accelerator, and they are often used in error-prone applications that are tolerant. There are studies that suggest several estimation methods. NNs exhibit parallelism and could be accelerated by special hardware [1]. There exist various quality measures such as pixel differences in an image, categorization of data and clusters, ranking accuracy, etc., and these can be subject to incorrect computations. Other quality measures include overall image quality index and validation. For many applications, there exist several performance measures that can be utilized to calculate quality reduction, such as k-means clustering accuracy and average centroid distance, which are being used as performance metrics [2]. Other areas include image processing, face detection, and search engines. Approximate computing always contributes to a variety of devices and components such as analytic models, CPUs and GPUs, simulators and inaccurate computation techniques, and eventually SRAM cells and cache memory.
General Aspects Related to the Field of Big Data and Big Data in Power Engineering
Published in Valentin A. Boicea, Energy Management, 2021
Another promising solution dedicated to the reduction of computer resources is the so-called approximate computing. This allows also a reduction in the consumed electrical energy [2]. This type of environment is used on a small synopsis of data instead of the entire data set. In most situations, when working with Big Data is highly probable to obtain approximate results and not exact values. That is because the Big Data set can contain outliers or values generated as a result of communication faults. Given also the fact that approximate computing is much more efficient in terms of processing time, one can say that it represents a compromise solution between speed and accuracy.
Low-power and high-speed approximate 4:2 compressors for image multiplication applications in CNFETs
Published in International Journal of Electronics, 2021
Danial Rostami, Mohammad Eshghi, Yavar Safaei Mehrabani
As discussed earlier, the main objective of approximate computing is to improve the hardware parameters of circuits, which subsequently lead to some quality penalties for output results. Therefore, to have a plausible and comprehensive arbitrage of the effectiveness of different circuits in approximate computing, we should take into account both hardware and quality metrics simultaneously. Since various parameters are involved in circuit and application levels, we present several figures of merit (FOM) that consider all of these parameters. FOM1 considers 1− NED as an accuracy metric and PDP as a well-established performance metric. FOM1 can be defined by Equation (13). Since the target application in this work is image multiplication, FOM2 and FOM3 are presented to provide a clear assessment of the suitability of different approximate designs. However, a metric was proposed in (Gupta et al., 2011) as PSNR2 × area × power. This metric completely ignores the delay of the circuits. It is notable that in addition to the PSNR, the SSIM is another important criterion for evaluating image quality that is not considered in this factor. Therefore, FOM2 considers delay, PSNR, and SSIM criteria (Equation (14)). Since the number of transistors is another important circuit evaluation figure of merit, this factor is taken into account in FOM3 (Equation (15)).
A new twelve-transistor approximate 4:2 compressor in CNTFET technology
Published in International Journal of Electronics, 2019
Samira Shirinabadi Farahani, Mohammad Reza Reshadinezhad
Approximate computing is a new major concern regarding efficient arithmetic power circuits’ design. A wide range of signal processing, multimedia, data mining, redundant data analytic applications are not sensitive to lack of accuracy. Approximate computing consists of general methods, and is run at hardware and software levels. The approximate algorithms are applied at software level. The approximate functional units like approximate adders, multipliers and dividers can be applied at hardware level. At hardware level approximation, altering the truth table, simplify logic complexity and truncation methods are applied the most frequently. In some cases, the circuit can switch between approximate and accurate mode to meet the runtime constraints. In some approximate designs, extra correction units are added to the main processing unit to compensate probable errors. In general approximate computing makes a trade-off between accuracy and power-area-delay efficiency (Akbari, Kamal, Afzali-Kusha, & Pedram, 2017; Golub, Jakobović, & Budin, 2000; Momeni, Han, Montuschi, & Lombardi, 2015; Monajati, Fakhraie, & Kabir, 2015; Zareei, Navi, & Keshavarziyan, 2018; Zervakis, Tsoumanis, Xydis, Soudris, & Pekmestzi, 2016).
Comparison and design of energy-efficient approximate multiplier schemes for image processing by CNTFET
Published in International Journal of Electronics, 2023
Elmira Tavakkoli, Shayan Shokri, Mahdi Aminian
Approximate computing is an emerging approach in the digital electronics, where a high degree of accuracy is not important. In many applications, such as image processing, accurate results are not necessary. For these applications, approximate circuits play an important role to achieve better performance in energy consumption (Momeni et al., 2015; Sabetzadeh et al., 2019). Multipliers are widely used in digital systems and arithmetic logic unit (ALU). Due to the reduction in complexity of multipliers hardware circuit and improvement of the overall systems performance, using approximate multipliers are an appropriate choice in energy-efficient systems (Zakian & Asli, 2020).