Power-efficient compute-in-memory pooling

Patent metadata

Field	Value
Publication number	US-2022012580-A1
Application number	US-202016922953-A
Country	US
Kind code	A1
Filing date	Jul 7, 2020
Priority date	Jul 7, 2020
Publication date	Jan 13, 2022
Grant date	—

Abstract

A multiply-and-accumulate (MAC) circuit having a plurality of compute-in-memory bitcells is configured to multiply a plurality of stored weight bits with a plurality of input bits to provide a MAC output voltage. A successive approximation analog-to-digital converter includes a capacitive-digital-to-analog-converter (CDAC) configured to subtract a bias voltage from the MAC output voltage to provide a CDAC output voltage. A comparator compares the CDAC output voltage to a fixed reference voltage.

First claim

We claim:

1 . A system for a machine learning application, comprising:
a first multiply-and-accumulate (MAC) circuit including a first plurality of compute-in-memory bitcells configured to multiply a first plurality of stored weights with an input vector to provide a first MAC output voltage; and an analog-to-digital converter configured to digitize the first MAC output voltage, the analog-to-digital converter including: a first capacitive digital-to-analog converter (CDAC) configured to subtract a bias voltage from the first MAC output voltage to provide a first CDAC output voltage; and a first comparator configured to compare the first CDAC output voltage to a reference voltage to provide a first comparator output signal.

2 . The system of claim 1 , wherein the reference voltage is approximately one-half of a power supply voltage.

3 . The system of claim 1 , wherein the bias voltage equals a difference between a threshold voltage for the first CDAC and the reference voltage, the analog-to-digital converter further comprising a control logic circuit configured to control the threshold voltage.

4 . The system of claim 3 , wherein the control logic circuit is a finite state machine.

5 . The system of claim 1 , further comprising:
a second MAC circuit including a second plurality of compute-in-memory bitcells configured to multiple a second plurality of stored weights with the input vector to provide a second MAC output voltage; and wherein the analog-to-digital converter further includes: a second CDAC configured to subtract the bias voltage from the second MAC output voltage to provide a second CDAC output voltage; a first logic gate configured to assert a local enable signal responsive to the first comparator output signal being false; a second comparator configured to respond to an assertion of the local enable signal to compare the second CDAC output voltage to the reference voltage to provide a second comparator output signal; and a second logic gate configured to provide a maximum pooling output signal for the first MAC circuit and for the second MAC circuit responsive to the first comparator output signal and to the second comparator output signal.

6 . The system of claim 5 , wherein the first logic gate is an AND gate configured to AND a complement of the first comparator output signal and a global enable signal to produce the local enable signal.

7 . The system of claim 5 , wherein the second logic gate is an OR gate configured to OR the first comparator output signal and the second comparator output signal to provide the maximum pooling output signal.

8 . A system for a machine learning application, comprising:
a first multiply-and-accumulate (MAC) circuit including a first plurality of compute-in-memory bitcells configured to multiply a first plurality of stored weights with an input vector to provide a first MAC output voltage; a second MAC circuit including a second plurality of compute-in-memory bitcells configured to multiply a second plurality of stored weights with the input vector to provide a second MAC output voltage; and an analog-to-digital converter configured to digitize either the first MAC output voltage or the second MAC output voltage, the analog-to-digital converter including: a first sampling switch; a second sampling switch; a first capacitive digital-to-analog converter (CDAC) configured to sample the first MAC output voltage through the first sampling switch to provide a sampled first MAC output voltage and to subtract a first bias voltage from the sampled first MAC output voltage to provide a first CDAC output voltage; a second capacitive digital-to-analog converter (CDAC) configured to sample the second MAC output voltage through the second sampling switch to provide a sampled second MAC output voltage and to subtract a second bias voltage from the sampled second MAC output voltage to provide a second CDAC output voltage; and a comparator configured to assert a comparator output signal responsive to the first CDAC output voltage being greater than the second CDAC output voltage.

9 . The system of claim 8 , further comprising:
a selection logic circuit configured to close the first sampling switch and to open the second sampling switch responsive to an assertion of the comparator output signal; and a control logic circuit configured to control the first bias voltage during a digitization of the sampled first MAC output voltage by the first CDAC.

10 . The system of claim 9 , wherein the comparator is further configured to compare the first CDAC output voltage to a reference voltage during the digitization of the sampled first MAC output voltage by the first CDAC.

11 . The system of claim 10 , wherein the reference voltage is one-half of a power supply voltage.

12 . A system for a machine learning application, comprising:
a multiply-and-accumulate (MAC) circuit including a plurality of compute-in-memory bitcells configured to multiply a plurality of stored weights with a first input vector to provide a first MAC output voltage and to multiply the plurality of stored weights with a second input vector to provide a second MAC output voltage; and an analog-to-digital converter configured to digitize either the first MAC output voltage or the second MAC output voltage, the analog-to-digital converter including: a capacitive digital-to-analog converter (CDAC) configured to subtract a bias voltage from the first MAC output voltage to provide a first CDAC output voltage and to subtract the bias voltage from the second MAC output voltage to provide a second CDAC output voltage; a comparator configured to compare the first CDAC output voltage to a reference voltage to provide a first comparator output signal responsive to a first assertion of a local enable signal; and a logic gate configured to perform a second assertion of the local enable signal responsive to the first comparator output signal being false, wherein the comparator is further configured to compare the second CDAC output voltage to the reference voltage to provide a second comparator output signal responsive to the second assertion of the local enable signal.

13 . The system of claim 12 , wherein the reference voltage is one-half of a power supply voltage.

14 . The system of claim 12 , wherein the logic gate is an AND gate.

15 . The system of claim 12 , wherein the logic gate comprises a pair of AND gates.

16 . The system of claim 12 , further comprising a delay circuit to delay the first comparator output signal.

17 . The system of claim 12 , further comprising a storage device configured to store the first comparator output signal and the second comparator output signal.

18 . A system for a machine learning application, comprising:
a multiply-and-accumulate (MAC) circuit including a plurality of compute-in-memory bitcells configured to multiply a plurality of stored weights with a first input vector to provide a first MAC output voltage and to multiply the plurality of stored weights with a second input vector to provide a second MAC output voltage; and an analog-to-digital converter configured to digitize either the first MAC output voltage or the second MAC output voltage, the analog-to-digital converter including: a first sampling switch; a second sampling switch; a first capacitive digital-to-analog converter (CDAC) configured to sample the first MAC output voltage through the first sampling switch to provide a sampled first MAC output voltage; and a second capacitive digital-to-analog converter

Assignees

• Qualcomm Inc

Inventors

• Srivastava Ankit

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06N3/065 Primary

  Analogue means · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G11C11/54

  using elements simulating biological cells, e.g. neuron · CPC title

• G11C11/4096

  Input/output [I/O] data management or control circuits, e.g. reading or writing circuits, I/O drivers or bit-line switches · CPC title