Mixed-signal programmable non-linear interface for resource-efficient multi-sensor analytics

Tremendous progress has been made in reducing ADC power-consumption, yet, in many portable always-awake and multi-sensor systems, the power consumption is dominated by digital backend processing [1] for feature-computation and classification. Recent Analog-to-Information based innovations (see Fig. 21.1.1) have attempted to alleviate this bottleneck by reducing the amount of data streaming into the digital domain: (a) by extracting sensory features in the analog domain [2] and digitizing these instead of the raw-data; and (b) by compressing the data through an analog non-linearity in order to reduce the required digitization word-length [3-5]. Yet, both approaches suffer from limited applicability and design reuse problems, due to (a) their need for highly application-specific building blocks which are not portable across different sensor-signals in multi-sensor platforms; and (b) their need for complex, performance-sensitive analog building blocks which are not portable across silicon technologies. Overcoming the above shortcomings, this work reports a highly programmable non-linear interface that synergistically combines a digitally-computed, application-tunable non-linearity with a 10b binary DAC in an iterative mixed-signal loop (see Fig. 21.1.1 bottom) to enable a compressive analog to non-linear digital transfer-curve. Such configurable non-linear transfer-curve has wide applicability in multi-sensor analytics for feature-extraction, signal-emphasis/-de-emphasis, signal-correction, etc. A 90nm CMOS proof-of-concept illustrates this versatility with 2 very different application mappings, demonstrating (a) Compressive classification: 2x improvement in rms-error for heart-beat classification from muscle noise corrupted ECG signals, along with 50% backend data reduction; and (b) Analog impairment correction : up-to 20dB distortion correction for analog impairments in the sensory chain.