TY - BOOK

T1 - KAVUAKA

T2 - A Low-Power Application-Specific Processor Architecture for Digital Hearing Aids

AU - Gerlach, Lukas Konrad

N1 - Doctoral thesis

PY - 2021

Y1 - 2021

N2 - The power consumption of digital hearing aids is very restricted due to their small physical size and the available hardware resources for signal processing are limited. However, there is a demand for more processing performance to make future hearing aids more useful and smarter. Future hearing aids should be able to detect, localize, and recognize target speakers in complex acoustic environments to further improve the speech intelligibility of the individual hearing aid user. Computationally intensive algorithms are required for this task. To maintain acceptable battery life, the hearing aid processing architecture must be highly optimized for extremely low-power consumption and high processing performance.The integration of application-specific instruction-set processors (ASIPs) into hearing aids enables a wide range of architectural customizations to meet the stringent power consumption and performance requirements. In this thesis, the application-specific hearing aid processor KAVUAKA is presented, which is customized and optimized with state-of-the-art hearing aid algorithms such as speaker localization, noise reduction, beamforming algorithms, and speech recognition. Specialized and application-specific instructions are designed and added to the baseline instruction set architecture (ISA). Among the major contributions are a multiply-accumulate (MAC) unit for real- and complex-valued numbers, architectures for power reduction during register accesses, co-processors and a low-latency audio interface. With the proposed MAC architecture, the KAVUAKA processor requires 16 % less cycles for the computation of a 128-point fast Fourier transform (FFT) compared to related programmable digital signal processors. The power consumption during register file accesses is decreased by 6 %to 17 % with isolation and by-pass techniques. The hardware-induced audio latency is 34 %lower compared to related audio interfaces for frame size of 64 samples.The final hearing aid system-on-chip (SoC) with four KAVUAKA processor cores and ten co-processors is integrated as an application-specific integrated circuit (ASIC) using a 40 nm low-power technology. The die size is 3.6 mm2. Each of the processors and co-processors contains individual customizations and hardware features with a varying datapath width between 24-bit to 64-bit. The core area of the 64-bit processor configuration is 0.134 mm2. The processors are organized in two clusters that share memory, an audio interface, co-processors and serial interfaces. The average power consumption at a clock speed of 10 MHz is 2.4 mW for SoC and 0.6 mW for the 64-bit processor.Case studies with four reference hearing aid algorithms are used to present and evaluate the proposed hardware architectures and optimizations. The program code for each processor and co-processor is generated and optimized with evolutionary algorithms for operation merging,instruction scheduling and register allocation. The KAVUAKA processor architecture is com-pared to related processor architectures in terms of processing performance, average power consumption, and silicon area requirements.

AB - The power consumption of digital hearing aids is very restricted due to their small physical size and the available hardware resources for signal processing are limited. However, there is a demand for more processing performance to make future hearing aids more useful and smarter. Future hearing aids should be able to detect, localize, and recognize target speakers in complex acoustic environments to further improve the speech intelligibility of the individual hearing aid user. Computationally intensive algorithms are required for this task. To maintain acceptable battery life, the hearing aid processing architecture must be highly optimized for extremely low-power consumption and high processing performance.The integration of application-specific instruction-set processors (ASIPs) into hearing aids enables a wide range of architectural customizations to meet the stringent power consumption and performance requirements. In this thesis, the application-specific hearing aid processor KAVUAKA is presented, which is customized and optimized with state-of-the-art hearing aid algorithms such as speaker localization, noise reduction, beamforming algorithms, and speech recognition. Specialized and application-specific instructions are designed and added to the baseline instruction set architecture (ISA). Among the major contributions are a multiply-accumulate (MAC) unit for real- and complex-valued numbers, architectures for power reduction during register accesses, co-processors and a low-latency audio interface. With the proposed MAC architecture, the KAVUAKA processor requires 16 % less cycles for the computation of a 128-point fast Fourier transform (FFT) compared to related programmable digital signal processors. The power consumption during register file accesses is decreased by 6 %to 17 % with isolation and by-pass techniques. The hardware-induced audio latency is 34 %lower compared to related audio interfaces for frame size of 64 samples.The final hearing aid system-on-chip (SoC) with four KAVUAKA processor cores and ten co-processors is integrated as an application-specific integrated circuit (ASIC) using a 40 nm low-power technology. The die size is 3.6 mm2. Each of the processors and co-processors contains individual customizations and hardware features with a varying datapath width between 24-bit to 64-bit. The core area of the 64-bit processor configuration is 0.134 mm2. The processors are organized in two clusters that share memory, an audio interface, co-processors and serial interfaces. The average power consumption at a clock speed of 10 MHz is 2.4 mW for SoC and 0.6 mW for the 64-bit processor.Case studies with four reference hearing aid algorithms are used to present and evaluate the proposed hardware architectures and optimizations. The program code for each processor and co-processor is generated and optimized with evolutionary algorithms for operation merging,instruction scheduling and register allocation. The KAVUAKA processor architecture is com-pared to related processor architectures in terms of processing performance, average power consumption, and silicon area requirements.

U2 - 10.15488/10576

DO - 10.15488/10576

M3 - Doctoral thesis

CY - Hannover

ER -