A Low-Power DSP Architecture for a Fully Implantable Cochlear Implant System-on-a-Chip.

Abstract

The National Science Foundation Wireless Integrated Microsystems (WIMS) Engineering Research Center at the University of Michigan developed Systems-on-a-Chip to achieve biomedical implant and environmental monitoring functionality in low-milliwatt power consumption and 1-2 cm3 volume. The focus of this work is implantable electronics for cochlear implants (CIs), surgically implanted devices that utilize existing nerve connections between the brain and inner-ear in cases where degradation of the sensory hair cells in the cochlea has occurred. In the absence of functioning hair cells, a CI processes sound information and stimulates the nderlying nerve cells with currents from implanted electrodes, enabling the patient to understand speech.

As the brain of the WIMS CI, the WIMS microcontroller unit (MCU) delivers the communication, signal processing, and storage capabilities required to satisfy the aggressive goals set forth. The 16-bit MCU implements a custom instruction set architecture focusing on power-efficient execution by providing separate data and address register windows, multi-word arithmetic, eight addressing modes, and interrupt and subroutine support. Along with 32KB of on-chip SRAM, a low-power 512-byte scratchpad memory is utilized by the WIMS custom compiler to obtain an average of 18% energy savings across benchmarks. A synthesizable dynamic frequency scaling circuit allows the chip to select a precision on-chip LC or ring oscillator, and perform clock scaling to minimize power dissipation; it provides glitch-free, software-controlled frequency shifting in 100ns, and dissipates only 480μW.

A highly flexible and expandable 16-channel Continuous Interleaved Sampling Digital Signal Processor (DSP) is included as an MCU peripheral component. Modes are included to process data, stimulate through electrodes, and allow experimental stimulation or processing. The entire WIMS MCU occupies 9.18mm2 and consumes only 1.79mW from 1.2V in DSP mode. This is the lowest reported consumption for a cochlear DSP.

Design methodologies were analyzed and a new top-down design flow is presented that encourages hardware and software co-design as well as cross-domain verification early in the design process. An O(n) technique for energy-per-instruction estimations both pre- and post-silicon is presented that achieves less than 4% error across benchmarks.

This dissertation advances low-power system design while providing an improvement in hearing recovery devices.