Neural network in a memory device

• US 9,430,735 B1
• Filed: 02/22/2013
• Issued: 08/30/2016
• Est. Priority Date: 02/23/2012
• Status: Active Grant

• Current Assignee: Micron Technology, Inc.
• Original Assignee: Micron Technology, Inc.
• Inventors: Vali, Tommaso, Eldredge, Kenneth J., Roohparvar, Frankie F., De Santis, Luca
• Primary Examiner(s): Chaki, Kakali
• Assistant Examiner(s): Coughlan, Peter
• Application Number:
  US13/774,553
• Time in Patent Office: 1,285 Days
• US Class Current: 1/1
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/063 using electronic means
• CPC Class Codes: G11C 16/0483 comprising cells having several storage transistors connected in series
• CPC Class Codes: G11C 16/08 Address circuits; Decoders; Word-line control circuits
• CPC Class Codes: G11C 16/26 Sensing or reading circuits; Data output circuits
• CPC Class Codes: G11C 7/1006 Data managing, e.g. manipulating data before writing or reading out, data bus switches or control circuits therefor
• CPC Class Codes: G11C 7/106 Data output latches

Abstract

Devices, systems and methods for operating a memory device facilitating a neural network in a memory device are disclosed. In at least one embodiment, the memory device is operated having a feed-ward neural network operating scheme. In at least one other embodiment, memory cells are operated to emulate a number of neural models to facilitate one or more neural network operating characteristics in the memory device.

Claims

1. A method of operating a memory device, the method comprising:

applying a first potential to a control gate of a first memory cell of a string of series-connected memory cells connected to a data line while activating each remaining memory cell of the string of series-connected memory cells;

generating a first output responsive to sensing the data line a first time while applying the first potential to the control gate of the first memory cell and while each remaining memory cell of the string of series-connected memory cells is activated;

applying a second potential to a control gate of a second memory cell of the string of series-connected memory cells while activating each remaining memory cell of the string of series-connected memory cells, wherein the second potential is generated as a function of the first output; and

sensing the data line a second time while applying the second potential to the control gate of the second memory cell and while each remaining memory cell of the string of series-connected memory cells is activated.

2. The method of claim 1, further comprising generating a second output responsive to sensing the data line the second time while applying the second potential to the control gate of the second memory cell and while each remaining memory cell of the string of series-connected memory cells is activated.

3. The method of claim 2, further comprising transferring the second output from the memory device.

4. The method of claim 1, further comprising:

applying a third potential to a control gate of a third memory cell of the string of series-connected memory cells while activating each remaining memory cell of the string of series-connected memory cells, wherein the third potential is generated as a function of the second output; and

sensing the data line a third time while applying the third potential to the control gate of the third memory cell and while each remaining memory cell of the string of series-connected memory cells is activated.

5. The method of claim 1, wherein the first potential comprises an input potential generated responsive to an input received at an interface of the memory device.

6. The method of claim 1, further comprising performing a programming operation on one or more memory cells of the string of series-connected memory cells prior to applying the first potential to the control gate of the first memory cell.

7. The method of claim 6, wherein performing the programming operation comprises adjusting a threshold voltage of one or more memory cells of the string of series-connected memory cells.

8. The method of claim 6, wherein performing the programming operation further comprises programming the one or more memory cells of the string of series-connected memory cells as SLC memory cells or MLC memory cells.

9. The method of claim 1, wherein sensing the data line the first time and sensing the data line the second time each comprise sensing a current on the data line while the first potential and the second potential are applied, respectively.

10. The method of claim 1, wherein sensing the data line the first time and sensing the data line the second time each comprises sensing a voltage on the data line while the first potential and the second potential are applied, respectively.

11. The method of claim 1, wherein the second potential comprises a logic high potential when the first output is greater than or equal to a particular threshold level, and wherein the second potential comprises a logic low potential when the first output is less than the particular threshold level.

12. A method of operating a memory device, the method comprising:

applying a first potential to a control gate of a first memory cell of a string of series-connected memory cells connected to a first data line and to a control gate of a first memory cell of a string of series-connected memory cells connected to a second data line while activating each remaining memory cell of the string of series-connected memory cells connected to the first data and each remaining memory cell of the string of series-connected memory cells connected to the second data line;

generating first and second outputs responsive to sensing the first and second data lines a first time while applying the first potential to the control gate of the first memory cell of the string of series-connected memory cells connected to the first data line and to the control gate of the first memory cell of the string of series-connected memory cells connected to the second data line, and while each remaining memory cell of the string of series-connected memory cells connected to the first data line and each remaining memory cell of the string of series-connected memory cells connected to the second data line is activated;

applying a second potential to a second memory cell of the string of series-connected memory cells connected to the first data line and to a second memory cell of the string of series-connected memory cells connected to the second data line while activating each remaining memory cell of the string of series-connected memory cells connected to the first data and each remaining memory cell of the string of series-connected memory cells connected to the second data line, wherein the second potential is generated as a function of at least one of the first and second outputs; and

generating third and fourth outputs responsive to sensing the first and second data lines a second time while applying the second potential to the control gate of the second memory cell of the string of series-connected memory cells connected to the first data line and to the control gate of the second memory cell of the string of series-connected memory cells connected to the second data line, and while each remaining memory cell of the string of series-connected memory cells connected to the first data line and each remaining memory cell of the string of series-connected memory cells connected to the second data line is activated.

13. The method of claim 12, further comprising:

applying a third potential to a control gate of a third memory cell of the string of series-connected memory cells connected to the first data line and to control gate of a third memory cell of the string of series-connected memory cells connected to the second data line while activating each remaining memory cell of the string of series-connected memory cells connected to the first data and each remaining memory cell of the string of series-connected memory cells connected to the second data line, wherein the third potential is generated as a function of at least one of the third and fourth outputs; and

generating fifth and sixth outputs responsive to sensing the first and second data lines a third time while applying the third potential to the control gate of the third memory cell of the string of series-connected memory cells connected to the first data line and to the control gate of the third memory cell of the string of series-connected memory cells connected to the second data line, and while each remaining memory cell of the string of series-connected memory cells connected to the first data line and each remaining memory cell of the string of series-connected memory cells connected to the second data line is activated.

14. The method of claim 13, further comprising:

applying a fourth potential to a fourth memory cell of the string of series-connected memory cells connected to the first data line and to a fourth memory cell of the string of series-connected memory cells connected to the second data line while activating each remaining memory cell of the string of series-connected memory cells connected to the first data and each remaining memory cell of the string of series-connected memory cells connected to the second data line, wherein the fourth potential is generated as a function of at least one of the fifth and sixth outputs; and

generating seventh and eighth outputs responsive to sensing the first and second data lines a fourth time while applying the fourth potential to the control gate of the fourth memory cell of the string of series-connected memory cells connected to the first data line and to the control gate of the fourth memory cell of the string of series-connected memory cells connected to the second data line, and while each remaining memory cell of the string of series-connected memory cells connected to the first data line and each remaining memory cell of the string of series-connected memory cells connected to the second data line is activated.

15. The method of claim 12, further comprising programming one or more memory cells of the strings of series-connected memory cells to one of a plurality of data states prior to applying the first potential to the control gate of the first memory cell of the string of series-connected memory cells connected to the first data line and to the control gate of the first memory cell of the string of series-connected memory cells connected to the second data line.

16. A method of operating a memory device, the method comprising:

applying a first potential to a control gate of a first memory cell of a first string of series-connected memory cells connected to a first data line while activating each remaining memory cell of the first string of series-connected memory cells;

performing a first sense operation on the first data line responsive to applying the first potential to the control gate of the first memory cell of the first string of series-connected memory cells;

generating a second potential, wherein the second potential comprises a potential level generated as a function of a first level sensed on the first data line during the first sense operation;

applying the second potential to a control gate of a first memory cell of a second string of series-connected memory cells connected to a second data line while activating each remaining memory cell of the second string of series-connected memory cells; and

performing a second sense operation on the second data line responsive to applying the second potential to the control gate of the first memory cell of the second string of series-connected memory cells.

17. The method of claim 16, further comprising:

generating a third potential, wherein the third potential comprises a potential level generated as a function of a second level sensed on the second data line during the second sense operation;

applying the third potential to a control gate of a second memory cell of the first string of series-connected memory cells connected to the first data line while activating each remaining memory cell of the first string of series-connected memory cells; and

performing a third sense operation on the first data line responsive to applying the third potential to the control gate of the second memory cell of the first string of series-connected memory cells.

18. The method of claim 17, further comprising:

generating a fourth potential, wherein the fourth potential comprises a potential level generated as a function of a third level sensed on the first data line during the third sense operation;

applying the fourth potential to a control gate of a second memory cell of the second string of series-connected memory cells connected to the second data line while activating each remaining memory cell of the second string of series-connected memory cells; and

performing a fourth sense operation on the second data line responsive to applying the fourth potential to the control gate of the second memory cell of the second string of series-connected memory cells.

19. The method of claim 16, wherein generating the second potential as a function of the first level sensed further comprises generating the second potential as a function of the first level sensed by accessing a table corresponding to the function, wherein an input to the table comprises the first level sensed and an output of the table comprises a value of the second potential to be generated.

20. The method of claim 19, further comprising storing the table in one or more memory locations of the memory device prior to generating the second potential.

21. The method of claim 20, wherein storing the table in the one or more memory locations further comprises storing the table where the one or more memory locations comprise volatile and/or non-volatile memory locations in the memory device.

22. A method of operating a memory device, the method comprising:

applying a first potential to a first access line coupled to a first memory cell of a first string of series-connected memory cells and to a first memory cell of a second string of series-connected memory cells while activating each remaining memory cell of the first string of series-connected memory cells and each remaining memory cell of the second string of series-connected memory cells, wherein the first string of series-connected memory cells is connected to a first data line and where the second string of series-connected memory cells is connected to a second data line;

applying a second potential to a second access line coupled to a first memory cell of a third string of series-connected memory cells and to a first memory cell of a fourth string of series-connected memory cells while activating each remaining memory cell of the third string of series-connected memory cells and each remaining memory cell of the fourth string of series-connected memory cells, wherein the third string of series-connected memory cells is connected to the first data line and where the fourth string of series-connected memory cells is connected to the second data line;

concurrently sensing the first data line and the second data line while applying the first potential and the second potential; and

generating a single output for a neuron model responsive to concurrently sensing the first and the second data lines.

23. The method of claim 22, wherein generating the single output further comprises generating the single output by adding current data corresponding to a current sensed on the first data line to current data corresponding to a current sensed on the second data line.

24. A method of operating a memory device as a neural network, the method comprising:

performing a first sense operation on a first plurality of memory cells connected to a first data line and comprising a first neuron model, and generating a first output, wherein each memory cell of the first plurality of memory cells is contained in a different string of series-connected memory cells of a first plurality of strings of series-connected memory cells and wherein each memory cell of the first plurality of strings of series-connected memory cells other than the first plurality of memory cells is activated while performing the first sense operation;

performing a second sense operation on a second plurality of memory cells connected to a second data line and comprising a second neuron model, and generating a second output, wherein each memory cell of the second plurality of memory cells is contained in a different string of series-connected memory cells of a second plurality of strings of series-connected memory cells and wherein each memory cell of the second plurality of strings of series-connected memory cells other than the second plurality of memory cells is activated while performing the second sense operation;

performing a third sense operation on a third plurality of memory cells connected to the first data line and comprising a third neuron model, wherein each memory cell of the third plurality of memory cells is contained in a different string of series-connected memory cells of the first plurality of strings of series-connected memory cells and wherein each memory cell of the first plurality of strings of series-connected memory cells other than the third plurality of memory cells is activated while performing the third sense operation; and

performing a fourth sense operation on a fourth plurality of memory cells connected to the second data line and comprising a fourth neuron model, wherein each memory cell of the fourth plurality of memory cells is contained in a different string of series-connected memory cells of the second plurality of strings of series-connected memory cells and wherein each memory cell of the second plurality of strings of series-connected memory cells other than the fourth plurality of memory cells is activated while performing the fourth sense operation;

wherein the third sense operation is performed by applying a first and a second potential to the third neuron model where the first potential and the second potential are generated as a function of the first output and the second output; and

wherein the fourth sense operation is performed by applying the first and the second potential to the fourth neuron model where the fourth sense operation is performed concurrently with the third sense operation.

25. A memory device, comprising:

a string of series-connected memory cells selectively connected to a data line; and

a controller, wherein the controller is configured to apply a first potential to a control gate of a first memory cell of the string of series-connected memory cells during a first sense operation on the data line while activating each remaining memory cell of the string of series-connected memory cells and while the string of series-connected memory cells is connected to the data line, and to apply a second potential to a control gate of a second memory cell of the string of series-connected memory cells during a second sense operation on the data line while activating each remaining memory cell of the string of series-connected memory cells and while the string of series-connected memory cells is connected to the data line, wherein the second potential comprises a potential generated as a function of an output generated responsive to sensing the data line during the first sense operation.

26. The memory device of claim 25, wherein the controller is further configured to apply a third potential to a control gate of a third memory cell of the series-connected string of memory cells during a third sense operation on the data line while activating each remaining memory cell of the string of series-connected memory cells and while the string of series-connected memory cells is connected to the data line, wherein the third potential is generated as a function of an output generated responsive to sensing the data line during the second sense operation.

27. The memory device of claim 26, wherein the controller is further configured to apply a fourth potential to a control gate of a fourth memory cell of the string of series-connected memory cells during a fourth sense operation on the data line while activating each remaining memory cell of the string of series-connected memory cells and while the string of series-connected memory cells is connected to the data line, wherein the fourth potential is generated as a function of an output generated responsive to sensing the data line during the third sense operation.

28. The memory device of claim 25, wherein the string of series-connected memory cells comprise charge storage memory cells.

29. The memory device of claim 25, wherein the string of series-connected memory cells comprise NAND configured strings of memory cells.

30. The memory device of claim 25, further comprising sense circuitry configured to sense a current on the data line during the sense operations.

31. The memory device of claim 30, wherein the generated output comprises a data value corresponding to a current sensed on the data line during the first sense operation.

32. The memory device of claim 31, wherein the second potential is generated as a function of the current sensed on the data line during the first sense operation.

33. The memory device of claim 32, wherein the function comprises one of a linear function or a non-linear function.

34. The memory device of claim 32, wherein the function comprises a sigmoid function.

35. The memory device of claim 32, wherein the function comprises a threshold function.

36. The memory device of claim 25, wherein the second potential is generated responsive to accessing a table.

37. The memory device of claim 25, wherein the controller is further configured to perform a program operation on one or more memory cells of the string of series-connected memory cells.

38. The memory device of claim 25, further comprising signal conditioning circuitry coupled to the data line.

39. The memory device of claim 25, further comprising one or more registers coupled to the data line.

40. The memory device of claim 25, wherein the controller comprises one or more access line drivers.

41. The memory device of claim 25, wherein the controller comprises one or more access line registers.

42. A system, comprising:

a communications channel;

a memory access device coupled to the communications channel and configured to generate memory device commands; and

a memory device coupled to the communications channel and configured to be responsive to the memory device commands, the memory device comprising;

a string of series-connected memory cells selectively connected to a data line; and

a controller, wherein the controller is configured to;

apply a first potential to a control gate of a first memory cell of the string of series-connected memory cells during a first sense operation on the data line while activating each remaining memory cell of the string of series-connected memory cells and while the string of series-connected memory cells is connected to the data line; and

apply a second potential to a control gate of a second memory cell of the string of series-connected memory cells during a second sense operation on the data line while activating each remaining memory cell of the string of series-connected memory cells and while the string of series-connected memory cells is connected to the data line, wherein the second potential comprises a potential generated as a function of an output generated responsive to sensing the data line during the first sense operation.

43. A memory device, comprising:

an array of memory cells comprising a plurality of NAND configured strings of series-connected memory cells; and

a controller, wherein the controller is configured to operate at least a portion of the array of memory cells as a neural network;

wherein the controller is further configured to operate a first plurality of memory cells of the array of memory cells as a first neuron model, wherein the first plurality of memory cells comprises at least two memory cells selectively connected to a particular data line, and wherein each of the at least two memory cells are contained in a different NAND configured string of series-connected memory cells each selectively connected to the particular data line; and

wherein the controller is further configured to sense the particular data line while each of the at least two memory cells receives a respective potential at its control gate to selectively activate that memory cell of the at least two memory cells depending upon its data state, and while each remaining memory cell of the different NAND configured strings of series-connected memory cells containing a respective one of the at least two memory cells is activated, and while the particular data line is connected to each of the different NAND configured strings of series-connected memory cells containing a respective one of the at least two memory cells.

44. The memory device of claim 43, wherein the controller is further configured to accept an initial input and apply the initial input to the control gates of the first plurality of memory cells and operate the first plurality of memory cells as the first neuron model responsive to the applied initial input.

45. The memory device of claim 43, wherein the controller is further configured to generate a first output responsive to operating the first plurality of memory cells as the first neuron model.

46. The memory device of claim 45, wherein the controller is further configured to operate a second plurality of memory cells of the array of memory cells as a second neuron model.

47. The memory device of claim 46, wherein the controller is further configured to operate the second plurality of memory cells as the second neuron model responsive to the generated first output.

48. The memory device of claim 47, wherein the controller is further configured to apply the generated first output to control gates of the second plurality of memory cells and to operate the second plurality of memory cells as the second neuron model responsive to the applied first output.

49. The memory device of claim 48, wherein the controller is further configured to generate a second output responsive to operating the second plurality of memory cells as the second neuron model.

50. The memory device of claim 49, wherein the second output comprises an output of the neural network.

51. The memory device of claim 49, wherein the first neuron model and the second neuron model comprise the neural network.

52. The memory device of claim 51, wherein the neural network comprises a feed-ward neural network.

53. The memory device of claim 46, wherein the second plurality of memory cells comprises at least two memory cells selectively connected to the particular data line, and wherein each of these at least two memory cells are contained in a different NAND configured string of series-connected memory cells.

54. The memory device of claim 46, wherein the second plurality of memory cells comprises at least two memory cells selectively connected to a different data line, and wherein the second plurality of memory cells are connected to receive a same set of inputs as the first plurality of memory cells.

55. A method of operating a memory device as a neural network, the method comprising:

stimulating a first plurality of memory cells of an array of memory cells comprising a plurality of strings of series-connected memory cells, wherein the first plurality of memory cells comprise a first neuron model, wherein the first plurality of memory cells comprises at least two memory cells selectively connected to a particular data line, and wherein each of the at least two memory cells are contained in a different string of a plurality of strings of series-connected memory cells;

activating each remaining memory cell of each of the strings of the plurality of strings of series-connected memory cells while stimulating the first plurality of memory cells and while each of the strings of the plurality of strings of series-connected memory cells is connected to the particular data line;

generating a first output responsive to stimulating the first plurality of memory cells while activating the remaining memory cells of the plurality of strings of series-connected memory cells;

stimulating a second plurality of memory cells of the memory cells of the array of memory cells wherein the second plurality of memory cells comprise a second neuron model, wherein the second plurality of memory cells comprises at least two memory cells selectively connected to the particular data line, and wherein each of these at least two memory cells are contained in a different string of the plurality of strings of series-connected memory cells; and

activating each remaining memory cell of each of the strings of the plurality of strings of series-connected memory cells while stimulating the second plurality of memory cells and while each of the strings of the plurality of strings of series-connected memory cells is connected to the particular data line;

wherein the second plurality of memory cells are stimulated by applying the first output to the second plurality of memory cells.

56. The method of claim 55, further comprising generating a second output responsive to stimulating the second plurality of memory cells with the first output while activating the remaining memory cells of the plurality of strings of series-connected memory cells.

57. The method of claim 55, further comprising receiving an input and wherein stimulating the first plurality of memory cells further comprises stimulating the first plurality of memory cells by applying the received input to respective control gates of one or more memory cells of the first plurality of memory cells.

58. A method of operating a memory device as a neural network, the method comprising:

stimulating a first neuron model, wherein the first neuron model comprises a first plurality of memory cells of an array of memory cells of the memory device, wherein the first plurality of memory cells comprises at least two memory cells selectively connected to a particular data line, and wherein each of the at least two memory cells are contained in a different string of a plurality of strings of series-connected memory cells of the array of memory cells;

activating each remaining memory cell of each of the strings of the plurality of strings of series-connected memory cells while stimulating the first neuron model and while each of the strings of the plurality of strings of series-connected memory cells is connected to the particular data line;

generating a first output responsive to stimulating the first neuron model while activating the remaining memory cells of the plurality of strings of series-connected memory cells;

stimulating a second neuron model, wherein the second neuron model comprises a second plurality of memory cells of the array of memory cells, wherein the second plurality of memory cells comprises at least two memory cells selectively connected to the particular data line, and wherein each of these at least two memory cells are contained in a different string of the plurality of strings of series-connected memory cells of the array of memory cells; and

activating each remaining memory cell of each of the strings of the plurality of strings of series-connected memory cells while stimulating the second neuron model and while each of the strings of the plurality of strings of series-connected memory cells is connected to the particular data line;

wherein the second neuron model is stimulated by applying the first output to the second neuron model.

59. The method of claim 58, wherein stimulating the second neuron model further comprises stimulating the second neuron model by concurrently applying the first output to respective control gates of the second plurality of memory cells.

60. The method of claim 58, further comprising generating a second output responsive to stimulating the second neuron model while activating the remaining memory cells of the plurality of strings of series-connected memory cells.

61. The method of claim 60, further comprising stimulating a third neuron model, wherein the third neuron model comprises a third plurality of memory cells of the array of memory cells, and where the third neuron model is stimulated by applying the second output to the third neuron model.

62. The method of claim 58, further comprising receiving an input prior to stimulating the first neuron model.

63. The method of claim 62, wherein stimulating the first neuron model further comprises stimulating the first neuron model by applying the received input to the first neuron model.

64. The method of claim 58, wherein the first neuron model and the second neuron model comprise the neural network.

65. The method of claim 64, further comprising programming one or more memory cells of the first plurality of memory cells and/or programming one or more memory cells of the second plurality of memory cells to one of a plurality of data states prior to stimulating the first neuron model, wherein the data states of the programmed memory cells define one or more operating characteristics of the neural network.

66. The method of claim 65, further comprising changing one or more programmed data states of one or more memory cells of the first and/or the second plurality of memory cells to change one or more operating characteristics of the neural network.

67. The method of claim 66, wherein changing one or more programmed data states further comprises changing one or more programmed data states by performing a programming operation and/or an erase operation on one or more memory cells of the first and/or the second plurality of memory cells.