ANALOG VERSUS DIGITAL Rahul Sarpeshkar Robert J Shillman

ANALOG VERSUS DIGITAL Rahul Sarpeshkar Robert J. Shillman Associate Professor MIT Electrical Engineering and Computer Science 6/10/04

A PRACTICAL APPLICATION 4 1 5 The charge from the electrode stimulation pulses is conducted to the spiral ganglion cell and activation occurs. 7 6 2 3 THE BIONIC EAR

OVERALL ARCHITECTURE OF ANALOG BIONIC EAR PROCESSOR 57 d. B Internal Dynamic Range Bandpass Filter Envelope Detector Log A-to-D Scanner FG 3329 AFE 77 d. B Overall Dynamic Range Bias Circuit AGC Bandpass Filter Envelope Detector Log A-to-D Scanner Visibility Followers Programmability DACs 7 -bit number 1 k. Hz Output bits

Bionic Ear Processor Chip 211 m. W total power consumption including a 96 m. W 80 d. B microphone front end. Ultra-Lowpower consumption unbeatable even at the end of Moore’s law 10 -20 years in the future because mike plus A-D-then-DSP approach will consume more than 211 m. W even if DSP power consumption is 0. Effective operation is at 5 m. W per MIP versus DSP’s 250 m. W per MIP.

SUMMARY • Digital: Robust in every signal (1 or 0) and in every device (just a switch) at the price of • • not exploiting available degrees of freedom for computation. Efficient because of a divide-and-conquer approach to computation. Most efficient at high precision. Analog: Not robust in every signal and in every device but robust in signal-flow path if attention is paid to reducing thermal noise, 1/f noise, power-supply noise, and device mismatch through clever design, feedback, and calibration. Efficient because you exploit all kinds of degrees of freedom in the signals and the devices for computation. Inefficient because it does not take a divide-and-conquer approach to computation. Most efficient at low precision. Hybrid: More efficient than either because it uses a divide-and-conquer approach to computation with constant signal restoration, while still exploiting analog degrees of freedom to maintain efficiency. There’s an optimal amount of analog processing before you make a signal-restoring digital decision and an optimal amount of division in the divide-and-conquer approach. Example of where delaying digitization and doing analog preprocessing gives you a huge win: The world’s lowest power programmable analog bionic ear processor. The power is so low that it can run a rechargeable 100 m. Ah battery for 30 years, and will be lower in power than an A-D-then-DSP implementation even at the end of Moore’s law. This is because its power is lower than that of a low power microphone and A/D alone. The processor is useful for ultra-low-power speech-recognition front ends as well. Other Examples: Neuron-inspired spike-based A-to-D converter with hybrid signal representation and Analog Memory