ANALOG VERSUS DIGITAL Rahul Sarpeshkar Robert J. Shillman Associate Professor MIT Electrical Engineering and Computer Science 11/19/03.

ANALOG VERSUS DIGITALANALOG VERSUS DIGITAL

Rahul SarpeshkarRobert J. Shillman Associate Professor

MITElectrical Engineering and Computer Science

11/19/03

HOW MUCH ANALOG DO YOU DO BEFORE YOU GO DIGITAL?HOW MUCH ANALOG DO YOU DO BEFORE YOU GO DIGITAL?

1

2

3

6

74 55

The charge from the electrode stimulation pulses is conducted to the spiral ganglion cell and activation occurs.

THE BIONIC EAR

A PRACTICAL APPLICATIONA PRACTICAL APPLICATION

ULTRA-LOW-POWER PROGRAMMABLE ANALOG ULTRA-LOW-POWER PROGRAMMABLE ANALOG BIONIC EAR PROCESSOR VIDEOBIONIC EAR PROCESSOR VIDEO

SUMMARYSUMMARY• 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 100mAh 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.