10 November 2017 in IEEE Journal of Solid-State Circuits
This paper presents a scalable inference accelerator called a deep-learning specific instruction-set processor (DSIP) to support various convolutional neural networks (CNNs). For CNNs requiring a large amount of computations and memory accesses, a programmable inference system called master-slave instruction set architecture (ISA) is newly proposed to achieve high flexibility, processing speed, and energy efficiency. The master is responsible for sending and receiving feature maps in order to deal with neural networks in a scalable way, and the slave performs CNN operations, such as multiply accumulate, max pooling, and activation functions, on the features received from the master. The master-slave ISA maximizes computation speed by overlapping the off-chip data transmission and the CNN operations, and reduces power consumption by performing the convolution incrementally to reuse input and partial-sum data as maximally as possible. An inference system can be configured by connecting multiple DSIPs in a form of either 1-D or 2-D chain structure in order to enhance computation speed further. To evaluate the proposed accelerator, a prototype chip is implemented and evaluated for AlexNet. Compared to the state-of-the-art accelerator, the DSIP-based system enhances the energy efficiency by 2.17x.
7 June 2018 in IEEE Transactions on Circuits and Systems I: Regular Papers
This paper presents a rescheduled dataflow of convolution and its hardware architecture that can enhance energy efficiency. For convolution involving a large amount of computations and memory accesses, previous accelerators employed parallel processing elements to meet real-time constraints. Though the previous approaches made a success in implementing complex convolution models, they load the same input features and filter weights from on-chip memories multiple times due to the iterative property of convolution operations, suffering from high energy consumption. To mitigate redundant memory accesses, a novel dataflow is proposed that computes convolution operations incrementally so as to reuse the loaded data as maximally as possible. In addition, several convolution accelerators supporting the rescheduled dataflow are investigated, and qualitative and quantitative analyses are performed to suggest a promising candidate for various convolution models. Simulation results show that the energy efficiency of the proposed accelerator outperforms that of the previous accelerator significantly.
30 March 2015 in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
This brief presents an energy-efficient architecture to extract mel-frequency cepstrum coefficients (MFCCs) for real-time speech recognition systems. Based on the algorithmic property of MFCC feature extraction, the architecture is designed with floating-point arithmetic units to cover a wide dynamic range with a small bit-width. Moreover, various operations required in the MFCC extraction are examined to optimize operational bit-width and lookup tables needed to compute nonlinear functions, such as trigonometric and logarithmic functions. In addition, the dataflow of MFCC extraction is tailored to minimize the computation time. As a result, the energy consumption is considerably reduced compared with previous MFCC extraction systems.
26 September 2013 in IEEE Transactions on Circuits and Systems II: Express Briefs
This brief presents a new area-efficient multimode encoder for long Bose-Chaudhuri-Hocquenghen codes. In the proposed multimode encoding architecture, several short linear-feedback shift registers (LFSRs) are cascaded in series to achieve the same functionality that a long LFSR has, and the output of a short LFSR is fed back to the input side to support multimode encoding. Whereas previous multimode architectures necessitate huge overhead due to preprocessing and postprocessing, the proposed architecture completely eliminates the overhead by exploiting an efficient transformation. Without sacrificing the latency, the proposed architecture reduces hardware complexity by up to 97.2% and 49.1% compared with the previous Chinese-remainder-theorem-based and weighted-summation-based multimode architectures, respectively.
25 February 2021 in IEEE Transactions on Circuits and Systems I: Regular Papers
This paper presents a convolution process and its hardware architecture for energy-efficient deep neural network (DNN) processing. A DNN in general consists of a number of convolutional layers, and the number of input features involved in the convolution of a shallow layer is larger than that of kernels. As the layer deepens, however, the number of input features decreases, while that of kernels increases. The previous convolution architectures developed for enhancing energy efficiency have tried to reduce the memory accesses by increasing the reuse of the data once accessed from the memory. However, redundant memory accesses are still required as the change in the numbers of data has not been considered. We propose a hybrid convolution process that selects either a kernel-stay or feature-stay process by taking into account the numbers of data, and a forwarding technique to further reduce the memory accesses needed to store and load partial sums. The proposed convolution process is effective in maximizing data reuse, leading to an energy-efficient hybrid convolution architecture. Compared to the state-of-the- art architectures, the proposed architecture enhances the energy efficiency by up to 2.38 times in a 65nm CMOS process.