Programming methods for neural network using non-volatile memory array
An artificial neural network device that utilizes one or more non-volatile memory arrays as the synapses. The synapses are configured to receive inputs and to generate therefrom outputs. Neurons are configured to receive the outputs. The synapses include a plurality of memory cells, wherein each of the memory cells includes spaced apart source and drain regions formed in a semiconductor substrate with a channel region extending there between, a floating gate disposed over and insulated from a first portion of the channel region and a non-floating gate disposed over and insulated from a second portion of the channel region. Each of the plurality of memory cells is configured to store a weight value corresponding to a number of electrons on the floating gate. The plurality of memory cells are configured to multiply the inputs by the stored weight values to generate the outputs. Various algorithms for tuning the memory cells to contain the correct weight values are disclosed.
1. A method of programming a non-volatile memory cell in a neural network, the method comprising
determining a first read current of the non-volatile memory cell;
when the first read current is greater than a target current, performing a programming tuning process to increase the number of electrons on a floating gate of the non-volatile memory cell;
determining a second read current of the non-volatile memory cell; and
when the second read current is less than the target current, and not more than the target current minus a delta value, performing an erasing tuning process to decrease the number of electrons on the floating gate of the non-volatile memory cell.
2. The method of claim 1 , wherein the programming tuning process comprises:
determining a programming voltage using a lookup table; and
applying the programming voltage to a terminal of the non-volatile memory cell.
3. The method of claim 2 , wherein the applying the programming voltage step is repeated one or more times.
4. The method of claim 1 , wherein the erasing tuning process comprises:
determining an erasing voltage using a lookup table; and
applying the erasing voltage to a terminal of the non-volatile memory cell.
5. The method of claim 4 , wherein the applying the erasing voltage step is repeated one or more times.
6. The method of claim 2 , wherein the erasing tuning process comprises:
determining an erasing voltage using a lookup table; and
applying the erasing voltage to a terminal of the non-volatile memory cell.
7. The method of claim 6 , wherein the applying the erasing voltage step is repeated one or more times.
8. The method of claim 3 , wherein the erasing tuning process comprises:
determining an erasing voltage using a lookup table; and
applying the erasing voltage to a terminal of the non-volatile memory cell.
9. The method of claim 8 , wherein the applying the erasing voltage step is repeated one or more times.
10. The method of claim 1 , wherein the non-volatile memory cell is a split-gate memory cell.
11. The method of claim 1 , wherein the non-volatile memory cell can be configured to store one of a plurality of weight level assignments that are evenly spaced.
12. The method of claim 1 , wherein the non-volatile memory cell can be configured to store one of a plurality of weight level assignments that are unevenly spaced.
13. The method of claim 1 , further comprising:
converting a set of weight digital bits into an analog voltage;
converting the analog voltage into an analog current;
determining a third read current of the non-volatile memory cell; and
comparing the analog current with the third read current to generate a signal indicating a weight stored in the non-volatile memory cell.
14. The method of claim 1 , further comprising:
converting a set of weight digital bits into an analog voltage;
determining a third read current of the non-volatile memory cell;
converting the third read current into a read voltage; and comparing the analog voltage with the read voltage to generate a signal indicating a weight stored in the non-volatile memory cell.
15. A method of programming a non-volatile memory cell in a neural network, the method comprising
determining a first read current of the non-volatile memory cell;
when the first read current is less than a target current, performing an erasing tuning process to decrease the number of electrons on a floating gate of the non-volatile memory cell;
determining a second read current of the non-volatile memory cell; and
when the second read current is greater than a target current, and not less than the target current plus a delta value, performing a programming tuning process to increase the number of electrons on the floating gate of the non-volatile memory cell.
16. The method of claim 15 , wherein the erasing tuning process comprises:
determining an erasing voltage using a lookup table; and
applying the erasing voltage to a terminal of the non-volatile memory cell.
17. The method of claim 16 , wherein the applying the erasing voltage step is repeated one or more times.
18. The method of claim 15 , wherein the programming tuning process comprises:
determining a programming voltage using a lookup table; and
applying the programming voltage to a terminal of the non-volatile memory cell.
19. The method of claim 18 , wherein the applying the programming voltage step is repeated one or more times.
20. The method of claim 16 , wherein the programming tuning process comprises:
determining a programming voltage using a lookup table; and
applying the programming voltage to a terminal of the non-volatile memory cell.
21. The method of claim 20 , wherein the applying the programming voltage step is repeated one or more times.
22. The method of claim 17 , wherein the programming tuning process comprises:
determining a programming voltage using a lookup table; and
applying the programming voltage to a terminal of the non-volatile memory cell.
23. The method of claim 22 , wherein the applying the programming voltage step is repeated one or more times.
24. The method of claim 15 , wherein the non-volatile memory cell is a split-gate memory cell.
25. The method of claim 15 , wherein the non-volatile memory cell can be configured to store one of a plurality of weight level assignments that are evenly spaced.
26. The method of claim 15 , wherein the non-volatile memory cell can be configured to store one of a plurality of weight level assignments that are unevenly spaced.
27. The method of claim 15 , further comprising:
converting a set of weight digital bits into an analog voltage;
converting the analog voltage into an analog current;
determining a third read current of the non-volatile memory cell; and
comparing the analog current with the third read current to generate a signal indicating a weight stored in the non-volatile memory cell.
28. The method of claim 15 , further comprising:
converting a set of weight digital bits into an analog voltage;
determining a third read current of the non-volatile memory cell;
converting the third read current into a read voltage; and
comparing the analog voltage with the read voltage to generate a signal indicating a weight stored in the non-volatile memory cell.
29. A method of programming a non-volatile memory cell in a neural network, the method comprising
determining a first read current of the non-volatile memory cell;
when the first read current is greater than a target current, performing a programming tuning process to increase the number of electrons on a floating gate of the non-volatile memory cell, the programming algorithm comprising:
determining a programming voltage from a lookup table;
applying the programming voltage to a terminal of the non-volatile memory cell;
determining a second read current; and
repeating the applying and determining a second read current steps until the second read current of the non-volatile memory cell is less than the target current and greater than the target current minus a delta value.
30. The method of claim 29 , wherein the non-volatile memory cell is a split-gate memory cell.
31. The method of claim 29 , wherein the non-volatile memory cell can be configured to store one of a plurality of weight level assignments that are evenly spaced.
32. The method of claim 29 , wherein the non-volatile memory cell can be configured to store one of a plurality of weight level assignments that are unevenly spaced.
33. The method of claim 29 , further comprising:
converting a set of weight digital bits into an analog voltage;
converting the analog voltage into an analog current;
determining a third read current of the non-volatile memory cell; and
comparing the analog current with the third read current to generate a signal indicating a weight stored in the non-volatile memory cell.
34. The method of claim 29 , further comprising:
converting a set of weight digital bits into an analog voltage;
determining a third read current of the non-volatile memory cell;
converting the third read current into a read voltage; and
comparing the analog voltage with the read voltage to generate a signal indicating a weight stored in the non-volatile memory cell.
35. A method of programming a non-volatile memory cell in a neural network, the method comprising
determining a first read current of the non-volatile memory cell;
when the first read current is less than a target current, performing an erasing tuning process to decrease the number of electrons on a floating gate of the non-volatile memory cell, the erasing algorithm comprising:
determining an erasing voltage from a lookup table;
applying the erasing voltage to a terminal of the non-volatile memory cell;
determining a second read current; and
repeating the applying and determining a second read current steps until the second read current of the non-volatile memory cell is greater than the target current and less than the target current plus a delta value.
36. The method of claim 35 , wherein the non-volatile memory cell is a split-gate memory cell.
37. The method of claim 35 , wherein the non-volatile memory cell can be configured to store one of a plurality of weight level assignments that are evenly spaced.
38. The method of claim 35 , wherein the non-volatile memory cell can be configured to store one of a plurality of weight level assignments that are unevenly spaced.
39. The method of claim 35 , further comprising:
converting a set of weight digital bits into an analog voltage;
converting the analog voltage into an analog current;
determining a third read current of the non-volatile memory cell; and
comparing the analog current with the third read current to generate a signal indicating a weight stored in the non-volatile memory cell.
40. The method of claim 35 , further comprising:
converting a set of weight digital bits into an analog voltage;
determining a third read current of the non-volatile memory cell;
converting the third read current into a read voltage; and
comparing the analog voltage with the read voltage to generate a signal indicating a weight stored in the non-volatile memory cell.