EERA-KWS: A 163 TOPS/W Always-on Keyword Spotting Accelerator in 28nm CMOS Using Binary Weight Network and Precision Self-Adaptive Approximate Computing

Abstract

This paper proposed an energy-efficient reconfigurable accelerator for keyword spotting (EERA-KWS) based on binary weight network (BWN) and fabricated in 28-nm CMOS technology. This keyword spotting system consists of two parts: the feature extraction based on melscale frequency cepstral coefficients (MFCC) and the keywords classification based on a BWN model, which is trained through the Google's Speech Commands database and deployed on our custom. To reduce the power consumption while maintaining the system recognition accuracy, we first optimize the MFCC implementation with approximate computing techniques, including Pre-emphasis coefficient transformation, rectangular Mel filtering, Framing and FFT optimization. Then, we propose a precision self-adaptive reconfigurable accelerator with digital-analog mixed approximate computing units to process the BWN efficiently. Based on the SNR prediction of background noise and post-detection of network output confidence, the BWN accelerator data path can be dynamically and adaptively reconfigured as 4, 8, or 16 bits. For the BWN accelerator, we proposed a time-delay based addition unit to process bit-wise approximate computing for the convolution layers and fully connected layers, and a LUT based unit for the activation layers. Implemented under TSMC 28 nm HPC+ process technology, the estimated power is 77.8 µW ∼ 115.9µW, the energy efficiency can achieve 163 TOPS/W, which is over 1.8× better than the state-of-the-art architecture.

Nanocomposite of cellulose acetate reinforced with nanocrystals modified chemically: Modification with bifunctional reagent

“…Changren Zhou et al have presented similar results in the literature. These authors' have observed that the elongations at break of NCPs were lower than matrix.…”

“…In general, the inclusion of CNCs caused great amendments on the viscoelastic behavior of NCs compared to matrix. The dynamic mechanical properties of the NCPs are strongly dependent on the filler orientation and concentration of nano reinforcements' as reported . The storage module of a NCP normally determines the effect of reinforcement in a polymer matrix .…”

“…Rarely percentages increase on mechanical resistance thus reaching those obtained in our work. Changren Zhou et al has investigated the effects of tensile strength and modulus of PLLA matrix reinforced with rod‐like chitin whiskers and surface modified with l ‐lactide to obtain grafted ones. The effects of the whiskers and surface modified whiskers on the mechanical properties of PLLA matrix were investigated.…”

A 14 μJ/Decision Keyword-Spotting Accelerator With In-SRAMComputing and On-Chip Learning for Customization

“…Since the VAD is for power savings, this could be integrated into this work as well. In this comparison, this work consumes less power than other works with similar model architecture [2], [5], [6] even though we process the entire raw data for the predicted results. [3], [4] use RNN as the model architecture, but [3] does not include feature extraction on the chip.…”

“…Guo et al [4] proposes hybrid digital circuits and 16 64x64 SRAM-based in-memory computing (IMC) macros with 3-bit ADC for recurrent neural network-based KWS. Liu et al [5] uses precision self-adaptive computing for a binary weight network to reduce power, and [6] uses mixed mode computing for low-power KWS. In summary, they reduce power consumption through recurrent models, quantized/binary neural network models, or voice activity detection.…”