Algorithms and circuitry for verifying a value stored during a programming operation of a non-volatile memory cell in an analog neural memory in deep learning artificial neural network
Various algorithms are disclosed for verifying the stored weight in a non-volatile memory cell in a neural network following a multilevel programming operation of the non-volatile memory cell by converting the stored weight into a plurality of digital output bits. Circuity, such as an adjustable reference current source, for implementing the algorithms are disclosed.
1. A verification method for a neural network for verifying the stored weight in a non-volatile memory cell following a multilevel programming operation of the non-volatile memory cell by converting the stored weight into a plurality of digital output bits, comprising:
comparing an output of the non-volatile memory cell against a reference value on a single reference line and generating a digital output bit of a first value if the memory cell output is greater than the reference value and generating a digital output bit of a second value if the memory cell output is less than the reference value; and
repeating the comparing step to generate each of the remaining digital output bits, wherein a reference value is selected based upon the digital output bit of the preceding comparing step.
2. The verification method of claim 1 , wherein the output of the non-volatile memory cell is a current or a voltage converted from the stored weight.
3. The verification method of claim 1 , wherein the first comparing step performed generates the most significant bit in the plurality of digital output bits.
4. The verification method of claim 1 , wherein the last comparing step performed generates the least significant bit in the plurality of digital output bits.
5. The verification method of claim 1 , wherein the non-volatile memory cell is a stacked-gate memory cell.
6. The verification method of claim 1 , wherein the non-volatile memory cell is a split-gate memory cell.
7. A verification method for verifying the stored weight in a non-volatile memory cell following a multilevel programming operation of the non-volatile memory cell for a neural network by converting the stored weight into a plurality of digital output bits, comprising:
comparing an output of the non-volatile memory cell against a reference value and generating a first digital output bit of a first value if the non-volatile memory cell output exceeds the reference value and generating a first digital output bit of a second value if the non-volatile memory cell output is less than the reference value; and
comparing the memory cell output against a second reference value if the first digital output bit has a first value and against a third reference value if the first digital output bit has a second value and performing one of:
generating a second digital output bit of a first value if the non-volatile memory cell output is greater than the second reference value and generating a second digital output bit of a second value if the stored voltage is less than the second reference value; and
generating a second digital output bit of a first value if the non-volatile memory cell output is greater than the third reference value and generating a second digital output bit of a second value if the non-volatile memory cell output is less than the third reference voltage.
8. The verification method of claim 7 , wherein the output of the memory cell is a current or a voltage converted from the stored weight.
9. The verification method of claim 7 , wherein the first comparing step performed generates the most significant bit in the plurality of digital output bits.
10. The verification method of claim 7 , wherein the non-volatile memory cell is a stacked-gate memory cell.
11. The verification method of claim 7 , wherein the non-volatile memory cell is a split-gate memory cell.
12. An adjustable reference current source for use during a verify operation following a programming operation of a non-volatile memory cell for a neural network, comprising:
an adjustable current source for providing an input voltage;
an array of devices for receiving the input voltage and control signals and generating an output current, wherein the control signals activate one or more devices in the array of devices in a thermometer coding fashion; and
a buffer mirror for receiving the output current and generating a reference voltage;
wherein the reference voltage varies in response to the adjustable voltage source and the control signals.
13. The adjustable reference current source of claim 12 , wherein the buffer mirror comprises an operational amplifier driving the mirror bias voltage.
14. The adjustable reference current source of claim 12 , wherein each device in the array of devices comprises a first NMOS transistor and a second NMOS transistor, wherein the source of the first NMOS transistor is coupled to the drain of the second NMOS transistor.
15. The adjustable reference current source of claim 12 , wherein the non-volatile memory cell is a split-gate memory cell.
16. The adjustable reference current source of claim 12 , wherein the non-volatile memory cell is a stacked-gate memory cell.