This letter proposes an unsigned approximate multiplier architecture segmented into three portions: the least significant portion that contributes least to the partial product (PP) is replaced with a new constant compensation term to improve hardware savings without sacrificing accuracy. The PPs in the middle portion are simplified using a new 4:2 approximate compressor, and the error due to approximation is compensated using a simple yet efficient error correction module. The most significant portion of the multiplier is implemented using exact logic as approximating it will results in a large error. Experimental results of 8-bit multiplier show that the power and power-delay products are reduced up to 47.7% and 55.2%, respectively, in comparison with the exact design and 36.9% and 39.5%, respectively, in comparison with the existing designs without significant compromise on accuracy.
Approximate computing is an evolving paradigm that aims to improve the power, speed, and area in neural network applications that can tolerate errors up to a specific limit. This letter proposes a new multiplier architecture based on the algorithm that adapts the approximate compressor from the existing and proposed compressors’ set to reduce error in the respective partial product columns. Further, the error due to the approximation in the proposed multiplier is corrected using a simple error-correcting module. Results prove that the power and power–delay product (PDP) of an 8-bit multiplier is improved by up to 39.9% and 43.6% compared with the exact multiplier and 27.5% and 23.9% compared to similar previous designs. The proposed multiplier is validated using image processing and neural network applications to prove its efficacy.
Approximate computing is a promising method for designing power-efficient computing systems. Many image and compression algorithms are inherently error-tolerant and can allow errors up to a specific limit. In such algorithms, savings in power can be achieved by approximating the data path units, such as a multiplier. This letter presents a novel decoder logic-based multiplier design with the intent to reduce the partial products generated. Thus, leading to a reduction in the hardware complexity and power consumption while maintaining a low error rate. Our proposed design in an 8-bit format which achieves 40.96% and 22.30% power reduction compared to the accurate and approximate multipliers. Comprehensive simulations are carried out on image sharpening and compression algorithms to prove that the proposed design obtains a better quality-effort tradeoff than the existing multipliers.
9 September 2021 in Journal of Circuits, Systems and Computers
The posit number system aims to be a drop-in replacement of the existing IEEE floating-point standard. Its properties-tapered precision and high dynamic range, allow a smaller size posit to almost match the performance of a much larger size floating-point in representing decimals. This becomes especially useful for performing error-tolerant tasks like deep learning inference computation where low latency and area are a priority. Recent research has found that the performance of deep neural network models saturates beyond a certain level of accuracy of multipliers used for convolutions. Therefore, the extra hardware cost of developing precise arithmetic circuits for such applications becomes an unnecessary overhead. This paper explores approximate posit multipliers in the convolutional layers of deep neural networks and attempts to find an ideal balance between hardware utilization and inference accuracy. Posit multiplication involves several steps, with the mantissa multiplication step utilizing maximum hardware resources. To mitigate this, a posit multiplier circuit using an approximate hybrid-radix Booth encoding for mantissa multiplication and techniques such as truncation and bit masking based on input regime size are proposed. In addition, a novel Booth encoding control scheme to prevent unnecessary bits from switching has been devised to reduce dynamic power dissipation. Compared to existing literature, these optimizations have contributed to a 23% decrease in power dissipation in the mantissa multiplication stage. Further, a novel area and energy-efficient decoder architecture have also been developed with an 11% reduction in dynamic power dissipation and area compared to existing decoders. Overall, the proposed < 16, 2 > posit multiplier offers a 14% reduction in the PDP over the existing approximate posit multiplier designs. The proposed < 16, 2 > multiplier also achieves over 90% accuracy in inferencing deep learning models such as ResNet20, VGG-19 and DenseNet.
The posit number system represents a binary numerical format devised primarily to mitigate the shortcomings of the IEEE 754 floating-point standard commonly used in modern computer arithmetic. Unlike IEEE 754, which utilizes a fixed-width representation, posits offer a variable-size encoding, dynamically adjusting the number of bits allocated based on the magnitude of the represented number. This adaptability enables posits to provide enhanced precision across a broader range of values, addressing issues like gradual underflow and the multiple representations of zero and NaN values inherent in IEEE 754. However, the dynamic bit allocation also poses significant challenges in decoding the posit number into its constituent fields and packing back the fields of the resultant posit once the arithmetic operations have been performed. For posits to become a viable alternative to floating-points in practical computing systems, the decoding and encoding overheads of posits need to be minimized. Hence, the aim of this paper is to develop energy-efficient hardware for posit decoding and encoding. The proposed < 16, 2 > posit decoders, Decoder A and Decoder B show an improvement of over 65% and 33% compared to literature in terms of energy requirements. Similarly, the proposed < 16, 2 > encoder circuit is over 52% more energy-efficient than existing encoder circuits.