IP Library › Granted Patent US 12,334,150
Granted Patent B2
US 12,334,150 · App. 18/188,978 · Granted Jun 17, 2025

Synaptic array for field-training-capable in-memory computing using non-volatile memory technologies

Inventor: Binh Quang Le (San Jose, CA)
Assignee: San Jose State University Research Foundation
G11C13/0069G11C11/54G11C13/0007G11C13/0026G11C13/0028G11C13/004G11C2013/0045G11C2213/72G11C2213/74G11C2213/79G11C2213/82
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12,334,150
App. No.
18/188,978
Granted
Jun 17, 2025
Kind
B2
Abstract

An apparatus and system are described to provide an in-memory computing non-volatile flash memory cell array used in a neural network. Each cell includes a Resistive RAM memory (RRAM) and a physical resistor formed from a high resistive material. The RRAM is programmed to either an on or off state in which the resistance is respectively significantly less or more than the resistor to permit the RRAM to act as a switch and allow for in-situ training. Multi-bit RRAM cells contain multiple RRAMs, each of which is connected to a resistor having a different resistance and read using the same input line. The resistors are formed from the same material as the resistor in the analog-to-digital converter used to read the array.

Claims (46)

1. An edge device comprising a Deep Neural Network (DNN), the DNN comprising:

a plurality of input nodes;

a plurality of output nodes; and

weights provided between the input nodes and the output nodes, the weights provided by a memory array, the memory array comprising a plurality of non-volatile memory cells, each non-volatile memory cell having a non-volatile memory coupled to a switch, each non-volatile memory cell coupled to:

a wordline coupled to the switch for selection of the non-volatile memory cell,

an input line to provide a current to the non-volatile memory cell through a resistor, and

an output line to which the non-volatile memory cell provides a current component of an output dependent on whether the non-volatile memory cell is in an on state, in which a resistance of the non-volatile memory is substantially less than the resistance of the resistor to which the non-volatile memory cell is connected, or an off state, in which the resistance of the non-volatile memory is substantially greater than the resistance of the resistor to which the non-volatile memory cell is connected.

2. The edge device of claim 1 , wherein:

the memory array further comprises a plurality of bitlines coupled with the non-volatile memory cells,

in a programming mode, each bitline is coupled to at least one of the non-volatile memory cells through a bitline select switch, and

in a read mode, each input line is coupled to at least one of the resistors through a read enable switch, each resistor coupled to a different one of the non-volatile memory cells.

3. A memory array comprising:

a plurality of non-volatile Random Access Memory (RAM) cells each comprising a first switch and a non-volatile memory coupled with the first switch, the first switch configured to be activated in each of a programming and read mode to respectively program and read the non-volatile memory; and

a plurality of Resistance-In-Memory RAM (RIMRAM) cells each comprising a second switch, a third switch, a unique one of the non-volatile memories, and a resistor coupled between the second switch and the unique one of the non-volatile memories, the second switch configured to be activated in the read mode to read the unique one of the non-volatile memories, the third switch configured to be activated in the programming mode to program the unique one of the non-volatile memories.

4. The memory array of claim 3 , wherein the resistors are formed from high sheet resistance material having a resistance of at least about 10 ohms/square.

5. The memory array of claim 4 , further comprising analog-to-digital converters (ADCs) coupled to the non-volatile memories, resistors of the ADCs formed from the high sheet resistance material to minimize variations caused by temperature and fabrication processes.

6. The memory array of claim 4 , wherein in the programming mode, each non-volatile memory is programmable between:

an on state in which a resistance of the non-volatile memory is substantially less than the resistance of the resistor to which the non-volatile memory cell is connected, and

an off state in which the resistance of the non-volatile memory is substantially greater than the resistance of the resistor to which the non-volatile memory cell is connected.

7. A memory array comprising:

a plurality of wordlines;

a plurality of input lines; and

an n-bit Resistance-In-Memory Random Access Memory (RIMRAM) cell, wherein n is an integer having a value of at least 2, coupled to associated wordlines of the plurality of wordlines and an input line of the plurality of input lines, the n-bit RIMRAM cell comprising a plurality of 1-bit RIMRAM cells, each 1-bit RIMRAM cell comprising:

a resistor that is coupled to the input line, and

a non-volatile memory cell comprising:

a switch, to which a respective wordline of the associated wordlines is coupled, and

a non-volatile memory coupled between the resistor and the switch.

8. The memory array of claim 7 , wherein the non-volatile memory comprises a Resistive RAM (RRAM).

9. The memory array of claim 7 , wherein the resistor is formed from high sheet resistance material having a resistance of at least about 10 ohms/square.

10. The memory array of claim 7 , wherein in a programming mode, each non-volatile memory is programmable between:

an on state in which a resistance of the non-volatile memory is substantially less than the resistance of the resistor to which the non-volatile memory cell is coupled, and

an off state in which the resistance of the non-volatile memory is substantially greater than the resistance of the resistor to which the non-volatile memory cell is coupled.

11. The memory array of claim 10 , further comprising a plurality of bitlines, the n-bit RIMRAM cell coupled to a bitline of the plurality of bitlines such that each non-volatile memory cell is coupled to the bitline through a different bitline select switch.

12. The memory array of claim 7 , wherein in a read mode, the input line is coupled to the resistors through a different read enable switch.

13. The memory array of claim 7 , wherein in a read mode, the input line is coupled to the resistors through a different read enable diode.

14. The memory array of claim 7 , wherein:

the input line is coupled to a digital-to-analog converter,

each non-volatile memory cell is coupled to an analog-to-digital converter, and

a resistor in the analog-to-digital converter is formed from identical material as the resistors in the n-bit RIMRAM cell.

15. The memory array of claim 7 , wherein:

the input line is coupled to different non-volatile memory cells of the n-bit RIMRAM cell through different resistors,

each of the different resistors has a different resistance, and

each of the different non-volatile memory cells of the n-bit RIMRAM cell is coupled to a same source line.

16. The memory array of claim 15 , wherein each of the different resistors has a resistance value that represents an integer multiple of a least resistance value among the resistors coupled to the input line.

17. The memory array of claim 15 , wherein each of the different resistors has a resistance value that represents 2 m of a least resistance value among the resistors coupled to the input line, where m is a non-negative integer.

18. The memory array of claim 7 , wherein the resistors of the n-bit RIMRAM cell and resistors of analog-to-digital converters to which the resistors of n-bit RIMRAM cell are coupled comprise a same material to minimize variations caused by temperature and fabrication processes.

Assignments (1)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded May 18, 2023
From: LE, BINH QUANG
To: SAN JOSE STATE UNIVERSITY RESEARCH FOUNDATION
Reel/Frame 063688/0122 →
Continuity (2)
Provisional Application 63325053 · Mar 29, 2022
Related Publication 20230317163A1 · Oct 5, 2023
References Cited (23)
US 9092736B2 · Aparin et al. · 2015 [cited by applicant]
US 11114158B1 · Ge · 2021 [cited by examiner]
US 11200948B1 · Ambrosi · 2021 [cited by examiner]
US 11861489B2 · Miao et al. · 2024 [cited by applicant]
US 20170040054A1 · Friedman · 2017 [cited by examiner]
US 20180322094A1 · Friedman · 2018 [cited by examiner]
US 20200034686A1 · Chiu et al. · 2020 [cited by applicant]
US 20220261624A1 · Song · 2022 [cited by applicant]
US 20230029509A1 · Lee · 2023 [cited by examiner]
EP 3459019 · 2019 [cited by applicant]
KR 101432197 · 2014 [cited by applicant]
WO 2019147522 · 2019 [cited by applicant]
WO 2020052342 · 2020 [cited by applicant]
WO 2023192795 · 2023 [cited by applicant]
“International Application Serial No. PCT US2023 064856, International Search Report mailed Jun. 20, 2023”, 4 pages. [cited by applicant]
“International Application Serial No. PCT US2023 064856, Written Opinion mailed Jun. 20, 2023”, 5 pages. [cited by applicant]
Jeong, Hongsik, “Memristor devices for neural networks”, Journal of Physics D: Applied Physics, Institute of Physics Publishing, Bristol, GB, vol. 52, No. 2, (Oct. 30, 2018), 28 pages. [cited by applicant]
Bankman, D., “RRAM-Based In-Memory Computing for Embedded Deep Neural Networks”, 2019 53rd Asilomar Conference on Signals, Systems, and Computers. IEEE, 2019., (Nov. 6, 2019), 5 pages. [cited by applicant]
Can, Li, “In-Memory Computing with Memristor Arrays”, 2018 IEEE International Memory Workshop (IMW). IEEE, 2018., (May 13, 2018), 4 pages. [cited by applicant]
Fick, L., “Analog In-Memory Subthreshold Deep Neural Network Accelerator”, 2017 IEEE Custom Integrated Circuits Conference (CICC). IEEE, 2017., (Apr. 30, 2017), 4 pages. [cited by applicant]
Hubara, Itay, “Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations”, Journal of Machine Learning Research 18.187 (2018): 1-30., (Apr. 2018), 30 pages. [cited by applicant]
Pozidis, H., “Enabling Technologies for Multilevel Phase-Change Memory”, European Phase Change and Ovonics Symposium, 2011, (2011), 8 pages. [cited by applicant]
“International Application Serial No. PCT/US2023/064856, International Preliminary Report on Patentability mailed Oct. 10, 2024”, 7 pgs. [cited by applicant]