IP Library › Granted Patent US 11,204,740
Granted Patent B2
US 11,204,740 · App. 16/695,509 · Granted Dec 21, 2021

Neuromorphic arithmetic device and operating method thereof

Inventors: Min-Hyung Cho (Daejeon, KR); Young-deuk Jeon (Sejong-si, KR); Ki Hyuk Park (Daejeon, KR); Joo Hyun Lee (Daejeon, KR)
Assignee: ELECTRONICS AND TELECOMMUNICATIONS RESEARCH INSTITUTE
G06F7/5443G06N3/063
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 11,204,740
App. No.
16/695,509
Granted
Dec 21, 2021
Kind
B2
Abstract

The neuromorphic arithmetic device performs a multiply-accumulate (MAC) calculation using a multiplier and an accumulator. The neuromorphic arithmetic device includes an offset accumulator configured to receive a plurality of offset data measured a plurality of times and accumulate the plurality of offset data, a bit extractor configured to obtain average offset data by extracting at least one first bit from the plurality of accumulated offset data, and a cumulative synapse array configured to accumulate a plurality of multiplication values generated by the multiplier and output a cumulative result of the plurality of multiplication values corrected according to the average offset data.

Claims (33)

1. A neuromorphic arithmetic device that performs a multiply-accumulate (MAC) calculation using a multiplier and an accumulator, the neuromorphic arithmetic device comprising:

an offset accumulator configured to receive a plurality of offset data measured a plurality of times and accumulate the plurality of offset data;

a bit extractor configured to obtain average offset data by extracting at least one first bit from the plurality of accumulated offset data; and

a cumulative synapse array configured to accumulate a plurality of multiplication values generated by the multiplier and output a cumulative result of the plurality of multiplication values corrected according to the average offset data.

2. The neuromorphic arithmetic device of claim 1 , wherein the cumulative synapse array comprises a first capacitor charged by a first voltage corresponding to a first multiplication value among the plurality of multiplication values.

3. The neuromorphic arithmetic device of claim 2 , wherein the cumulative synapse array further comprises a second capacitor charged by a second voltage corresponding to a second multiplication value among the plurality of multiplication values,

wherein the first and second capacitors are simultaneously charged.

4. The neuromorphic arithmetic device of claim 2 , further comprising a third capacitor charged by a third voltage corresponding to the first bit of the average offset data.

5. The neuromorphic arithmetic device of claim 4 , further comprising a fourth capacitor charged by a fourth voltage corresponding to the cumulative result of the plurality of multiplication values corrected according to the average offset data through connection of the first and third capacitors.

6. The neuromorphic arithmetic device of claim 5 , further comprising:

a first switch located between the first capacitor and the fourth capacitor;

a second switch located between the third capacitor and the fourth capacitor; and

an offset controller configured to control the first and second switches to connect the fourth capacitor with the first and third capacitors after the first and third capacitors are charged.

7. The neuromorphic arithmetic device of claim 1 , wherein the bit extractor further extracts at least one second bit from the plurality of accumulated offset data, and

the neuromorphic arithmetic device comprises:

a fifth capacitor charged by a fifth voltage corresponding to the at least one first bit; and

a sixth capacitor charged by a sixth voltage corresponding to the at least one second bit,

wherein a capacitance of the fifth capacitor is different from a capacitance of the sixth capacitor.

8. The neuromorphic arithmetic device of claim 1 , wherein the plurality of times is 2 k , where k is an integer of 1 or more.

9. The neuromorphic arithmetic device of claim 8 , wherein when the bit extractor extracts a plurality of bits from the plurality of accumulated offset data, the plurality of bits are continuous.

10. An operating method of a neuromorphic arithmetic device including an offset accumulator, a bit extractor, a multiplier, and a cumulative synapse array to perform a multiply-accumulate (MAC) calculation, the operating method comprising:

measuring a plurality of offset data by a plurality of times;

accumulating, by the offset accumulator, the plurality of offset data;

obtaining, by the bit extractor, an average offset data by extracting at least one first bit from the plurality of accumulated offset data;

calculating, by the multiplier, a plurality of multiplication values between a plurality of feature data and a plurality of weight data;

accumulating, by the cumulative synapse array, the plurality of multiplication values; and

correcting, by the cumulative synapse array, a cumulative result of the plurality of multiplication values according to the average offset data.

11. The operating method of claim 10 , wherein the neuromorphic arithmetic device comprises:

a plurality of accumulation capacitors configured to store the plurality of multiplication values, respectively; and

a correction capacitor configured to store the cumulative result;

wherein the accumulating of the plurality of multiplication values comprises simultaneously connecting the plurality of accumulation capacitors with the correction capacitor.

12. The operating method of claim 11 , wherein the neuromorphic arithmetic device further comprises a plurality of offset capacitors configured to store a plurality of bits of the average offset data, respectively,

wherein the correcting of the cumulative result of the plurality of multiplication values according to the average offset data further comprises connecting the plurality of offset capacitors with the correction capacitor.

Assignments (1)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Nov 26, 2019
From: CHO, MIN-HYUNG; JEON, YOUNG-DEUK; PARK, KI HYUK; LEE, JOO HYUN
To: ELECTRONICS AND TELECOMMUNICATIONS RESEARCH INSTITUTE
Reel/Frame 051116/0135 →
Priority Claims (2)
KR 10-2018-0154817 · Dec 4, 2018 · national
KR 10-2019-0103396 · Aug 23, 2019 · national
Continuity (1)
Related Publication 20200174751A1 · Jun 4, 2020