Nonvolatile tunable capacitive processing unit

A method of performing a multiply-accumulate (MAC) operation for a neural network (NN), the method comprising:

• providing one or more circuit configurations, wherein the one or more circuit configurations comprise:
• a charging transistor comprising:
• a first charging transistor terminal connected to a voltage line;
• and
• a second charging transistor terminal connected to a first capacitor terminal of a tunable capacitor, wherein the tunable capacitor comprises a second capacitor terminal connected to a ground terminal;
• a discharging transistor comprising:
• a first discharging terminal connected to the second charging transistor terminal and the first capacitor terminal;
• and
• a second discharging terminal connected to a current line;
• providing a predefined amount of heat to the tunable capacitor for a predefined amount of time, wherein the tunable capacitor has a first capacitance value and the predefined amount of heat is sufficient to change the first capacitance value to a second capacitance value;
• charging the tunable capacitor for a first predefined amount of time by applying an input voltage to the tunable capacitor via the charging transistor in a closed state and the discharging transistor in an open state;
• discharging the tunable capacitor for a second predefined amount of time by stopping the input voltage to the tunable capacitor via the charging transistor in the open state and the discharging transistor in the closed state, wherein charges flow from the tunable capacitor through the discharging transistor to an integrator circuit;
• determining a total charge of the one or more circuit configurations comprising the tunable capacitor by integrating the charges of the one or more circuit configurations through the discharging transistor using the integrator circuit;
• and
• generating a result for the MAC operation based, at least in part, on the total charge of the one or more circuit configurations.