IP Library Granted Patent US 12,632,220
Granted Patent B2
US 12,632,220 · App. 17/542,833 · Granted May 19, 2026

Apparatus and method with multiply-accumulate operation

Inventors: Hyungwoo Lee (Seoul, KR); Seungchul Jung (Suwon-si, KR); Sang Joon Kim (Hwaseong-si, KR); Sungmeen Myung (Suwon-si, KR)
Assignee: Samsung Electronics Co., Ltd.
G06F7/5443G06G7/16G06F2207/4824
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 12,632,220
App. No.
17/542,833
Granted
May 19, 2026
Kind
B2
Abstract

A multiply-accumulate (MAC) computation circuit includes: a source bit cell block configured to determine a MAC operation result of an input signal based on a plurality of source bit cells; a replica bit cell block comprising a plurality of replica bit cells corresponding to the plurality of source bit cells; and a readout circuit configured to read out a digital value of the MAC operation result using the replica bit cell block.

Claims (74)

1 . A multiply-accumulate (MAC) computation circuit comprising:

a source bit cell block configured to determine a MAC operation result of an input signal based on a plurality of source bit cells;

a replica bit cell block comprising a plurality of replica bit cells corresponding to the plurality of source bit cells; and

a readout circuit configured to receive the MAC operation result, adjust a state of at least a portion of the plurality of replica bit cells based on the MAC operation result, and read out a digital value of the MAC operation result using the replica bit cell block.

2 . The MAC computation circuit of claim 1 , wherein

the plurality of source bit cells are configured to determine resistance values indicating respective partial operation results,

a composite resistance value of the source bit cell block indicates the MAC operation result corresponding to a sum of the partial operation results, and

the readout circuit is configured to match a composite resistance value of the replica bit cell block to the composite resistance value of the source bit cell block by adjusting a resistance value of at least the portion of the plurality of replica bit cells.

3 . The MAC computation circuit of claim 2 , wherein the matched composite resistance value of the replica bit cell block indicates the digital value of the MAC operation result.

4 . The MAC computation circuit of claim 1 , wherein the readout circuit comprises a comparator configured to compare the source bit cell block and the replica bit cell block, and is configured to adjust the state of at least the portion of the plurality of replica bit cells based on a comparison result of the comparator.

5 . The MAC computation circuit of claim 4 , wherein

the replica bit cell block is connected in series to the source bit cell block, and

the comparator is configured to compare a composite resistance value of the source bit cell block and a composite resistance value of the replica bit cell block based on a voltage applied to the source bit cell block.

6 . The MAC computation circuit of claim 1 , wherein the readout circuit is configured to adjust the state of at least the portion of the plurality of replica bit cells using a binary search.

7 . The MAC computation circuit of claim 1 , wherein the readout circuit is configured to:

determine states of bit cells of a first group corresponding to ½ of the replica bit cell block based on a result of a comparison between the source bit cell block and the replica bit cell block in a first stage; and

determine states of bit cells of a second group corresponding to ½ of the remaining ½ of the replica bit cell block based on a result of a comparison between the source bit cell block and the replica bit cell block in a second stage.

8 . The MAC computation circuit of claim 7 , wherein

a first bit of the digital value of the MAC operation result is determined corresponding to the states of the bit cells of the first group, and

a second bit of the digital value of the MAC operation result is determined corresponding to the states of the bit cells of the second group.

9 . The MAC computation circuit of claim 1 , wherein

the replica bit cell block comprises a plurality of sub-blocks corresponding to different reference resistance values, and

the readout circuit is configured to detect a sub-block corresponding to a composite resistance value of the source bit cell block among the plurality of sub-blocks and read out the digital value of the MAC operation result using the detected sub-block.

10 . The MAC computation circuit of claim 1 , further comprising:

a plurality of source bit cell blocks including the source bit cell block;

a plurality of replica bit cell blocks including the replica bit cell block; and

a plurality of readout circuits including the readout circuit.

11 . An electronic apparatus comprising:

the MAC computation circuit of claim 1 ; and

one or more processors configured to generate a recognition result of an input corresponding to the input signal, based on the digital value.

12 . A multiply-accumulate (MAC) computation circuit comprising:

a plurality of local source bit cell blocks configured to each perform a MAC operation on an input signal based on a plurality of source bit cells;

a replica bit cell block comprising a plurality of replica bit cells corresponding to the plurality of source bit cells; and

a readout circuit configured to receive individual MAC operation results of the plurality of local source bit cell blocks, adjust a state of at least a portion of the plurality of replica bit cells based on the individual MAC operation results of the plurality of local source bit cell blocks, and read out digital values of the individual MAC operation results of the plurality of local source bit cell blocks using the replica bit cell block.

13 . The MAC computation circuit of claim 12 , wherein the plurality of local source bit cell blocks share the replica bit cell block and the readout circuit.

14 . The MAC computation circuit of claim 12 , wherein the readout circuit is configured to sequentially read out the individual MAC operation results of the plurality of local source bit cell blocks.

15 . The MAC computation circuit of claim 12 , wherein the readout circuit is configured to determine the digital values of the individual MAC operation results by adjusting the state of at least the portion of the plurality of replica bit cells.

16 . The MAC computation circuit of claim 12 , wherein

the plurality of source bit cells are configured to determine resistance values indicating respective partial operation results,

a composite resistance value of the local source bit cell blocks indicates the individual MAC operation results corresponding to a sum of the partial operation results, and

the readout circuit is configured to match a composite resistance value of the replica bit cell block to the composite resistance value of the local source bit cell blocks by adjusting a resistance value of at least the portion of the plurality of replica bit cells.

17 . The MAC computation circuit of claim 12 , wherein the readout circuit comprises a comparator configured to compare the plurality of local source bit cell blocks and the replica bit cell block, and is configured to adjust the state of at least the portion of the plurality of replica bit cells based on a comparison result of the comparator.

18 . The MAC computation circuit of claim 17 , wherein

the replica bit cell block is connected in series to a target bit cell block among the plurality of local source bit cell blocks, and

the comparator is configured to compare a composite resistance value of the target bit cell block and a composite resistance value of the replica bit cell block based on a voltage applied to the local source bit cell blocks.

19 . The MAC computation circuit of claim 12 , wherein the readout circuit is configured to adjust the state of at least the portion of the plurality of replica bit cells using a binary search.

20 . The MAC computation circuit of claim 12 , wherein

the plurality of local source bit cell blocks comprise a first local source bit cell block configured to determine a first individual MAC operation result based on the input signal,

the replica bit cell block comprises a plurality of sub-blocks corresponding to different reference resistance values, and

the readout circuit is configured to detect a sub-block corresponding to a composite resistance value of the first local source bit cell block among the plurality of sub-blocks and read out a first digital value of the first individual MAC operation result using the detected sub-block.

21 . A multiply-accumulate (MAC) computation method comprising:

determining a multiply-accumulate (MAC) operation result of an input signal based on a plurality of source bit cells;

adjusting, by a readout circuit, a state of at least a portion of a plurality of replica bit cells corresponding to the plurality of source bit cells, based on the MAC operation result; and

reading out a digital value of the MAC operation result using a replica bit cell block comprising the plurality of replica bit cells.

22 . A multiply-accumulate (MAC) computation method comprising:

controlling a replica bit cell block by setting, for each of groups of the replica bit cell block, replica bit cells of the group to have either one of a first resistance value and a second resistance value, based on a composite resistance value of a source bit cell block, wherein the composite resistance value represents a MAC operation result of an input signal; and

reading out a digital value of the MAC operation result of the source bit cell block by comparing, at a readout circuit, the composite resistance value of the source bit cell block and composite resistance values of the replica bit cell block generated by the controlling.

23 . The method of claim 22 , wherein, in a first stage,

the controlling comprises setting each of replica bit cells of a first group of the groups to have the first resistance value and each of replica bit cells of remaining groups of the groups to have the second resistance value, and

the reading out comprises determining a first bit of the digital value by comparing the composite resistance value of the source bit cell block and a composite resistance value of the replica bit cell block generated by the controlling in the first stage.

24 . The method of claim 23 , wherein the determining of the first bit comprises determining the first bit to be a first value in response to the composite resistance value of the source bit cell block being greater than or equal to the composite resistance value of the replica bit cell block generated by the controlling in the first stage.

25 . The method of claim 23 , wherein

the determining of the first bit comprises determining the value of the first bit to be a second value in response to the composite resistance value of the source bit cell block being less than the composite resistance value of the replica bit cell block generated by the controlling in the first stage, and

the first value is greater than the second value.

26 . The method of claim 23 , wherein the first bit is a most significant bit (MSB) of the digital value.

27 . The method of claim 22 , wherein a maximum value of the composite resistance value of the source bit cell block is twice a maximum value of the composite resistance value of the replica bit cell block generated by the controlling in the first stage.

28 . The method of claim 22 , wherein a number of the replica bit cells included in the first group is twice a number of replica bit cells included in a second group of the groups.

29 . An electronic apparatus comprising:

a multiply-accumulate (MAC) computation circuit comprising:

a source bit cell block configured to determine a MAC operation result of an input signal based on a plurality of source bit cells;

a replica bit cell block comprising a plurality of replica bit cells corresponding to the plurality of source bit cells; and

a readout circuit configured to receive the MAC operation result, adjust a state of at least a portion of the plurality of replica bit cells based on the MAC operation result, and read out a digital value of the MAC operation result using the replica bit cell block; and

one or more processors configured to generate a recognition result of an input corresponding to the input signal, based on the digital value.

30 . The apparatus of claim 29 , wherein the recognition result corresponds to any one or any combination of any two or more of object classification, object recognition, speech recognition, and image recognition.

Assignments (1)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Dec 6, 2021
From: LEE, HYUNGWOO; JUNG, SEUNGCHUL; KIM, SANG JOON; MYUNG, SUNGMEEN
To: SAMSUNG ELECTRONICS CO., LTD.
Reel/Frame 058307/0142 →
Priority Claims (1)
KR 10-2021-0061753 · May 13, 2021 · national
Continuity (1)
Related Publication 20220365752A1 · Nov 17, 2022
References Cited (25)
US 11226763B2 · Le Gallo-Bourdeau · 2022 [cited by examiner]
US 12327591B2 · Deshpande · 2025 [cited by examiner]
US 20050144210A1 · Simkins et al. · 2005 [cited by applicant]
US 20110320768A1 · Pell et al. · 2011 [cited by applicant]
US 20140172937A1 · Linderman · 2014 [cited by examiner]
US 20150170025A1 · Wu · 2015 [cited by examiner]
US 20180095748A1 · Buchanan et al. · 2018 [cited by applicant]
US 20190102170A1 · Chen et al. · 2019 [cited by applicant]
US 20190108193A1 · Buchanan · 2019 [cited by applicant]
US 20190132602A1 · Halawani et al. · 2019 [cited by applicant]
US 20200026989A1 · Zhang et al. · 2020 [cited by applicant]
US 20200210369A1 · Song · 2020 [cited by examiner]
US 20200303003A1 · Yanagisawa et al. · 2020 [cited by applicant]
US 20200356344A1 · Cruz-Albrecht et al. · 2020 [cited by applicant]
US 20210110251A1 · Gattani et al. · 2021 [cited by applicant]
US 20210366542A1 · Lee et al. · 2021 [cited by applicant]
KR 1020180036587A · 2018 [cited by applicant]
KR 1020200076571A · 2020 [cited by applicant]
KR 1020210000313A · 2021 [cited by applicant]
KR 1020210028063A · 2021 [cited by applicant]
KR 1020210144417A · 2021 [cited by applicant]
Extended European Search Report Issued on Jul. 4, 2022, in counterpart European Patent Application No. 22151759.2 (8 Pages). [cited by applicant]
Dong, Qing et al. “A 351TOPS/W and 372.4GOPS Compute-in-Memory SRAM Macro in 7nm FinFET CMOS for Machine-Learning Applications” [cited by applicant]
E. Sinangil, Mahmut et al. “A 7-nm Compute-in-Memory SRAM Macro Supporting Multi-Bit Input, Weight and Output and Achieving 351 TOPS/W and 372.4 Gops” IEEE Journal of Solid-State Circuits, vol. 56, No. Jan. 1, 2021 (11 … [cited by applicant]
Korean Office Action Issued on Aug. 26, 2025, in Counterpart Korean Patent Application No. 10-2021-0061753 (6 Pages in English, 8 Pages in Korean). [cited by applicant]