# US 2022/0044719 A1 · Probabilistic in-memory computing

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2022/0044719 A1
- Document ID: US20220044719A1
- Status: Published application
- Kind: A1
- Application number: 17/508,818
- Filed: 2021-10-22
- Published: 2022-02-10
- Earliest claimed: 2021-10-22
- Assignees: Intel Corporation
- Inventors: Hai Li; Dmitri E. Nikonov; Punyashloka Debashis; Ian A. Young; Mahesh Subedar; Omesh Tickoo
- Web page: https://patentgrep.com/patent/US20220044719A1

## CPC

- G11C 11/1673: {Reading or sensing circuits or methods}
- G06F 7/5443: {Sum of products (for applications thereof, see the relevant places, e.g. G06F17/10, H03H17/00)}
- G06N 3/045: Combinations of networks
- G06N 3/047: Probabilistic or stochastic networks
- G06N 3/0499: Feedforward networks
- G06N 3/065: Analogue means
- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N 7/01: Probabilistic graphical models, e.g. probabilistic networks
- G11C 11/1659: {Cell access}
- G11C 11/1675: {Writing or programming circuits or methods}
- G11C 11/1697: {Power supply circuits}
- G11C 11/18: using Hall-effect devices
- G11C 11/223: {using MOS with ferroelectric gate insulating film}
- G11C 11/2259: {Cell access}
- G11C 11/2273: {Reading or sensing circuits or methods}
- G11C 11/2275: {Writing or programming circuits or methods}
- G11C 11/54: using elements simulating biological cells, e.g. neuron
- G11C 11/5607: {using magnetic storage elements}
- G11C 11/5657: {using ferroelectric storage elements}
- G11C: STATIC STORES (semiconductor memory devices H10B)
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

Embodiments of the present disclosure are directed toward probabilistic in-memory computing configurations and arrangements, and configurations of probabilistic bit devices (p-bits) for probabilistic in-memory computing. concept with emerging. A probabilistic in-memory computing device includes an array of p-bits, where each p-bit is disposed at or near horizontal and vertical wires. Each p-bit is a time-varying resistor that has a time-varying resistance, which follows a desired probability distribution. The time-varying resistance of each p-bit represents a weight in a weight matrix of a stochastic neural network. During operation, an input voltage is applied to the horizontal wires to control the current through each p-bit. The currents are accumulated in the vertical wires thereby performing respective multiply-and-accumulative (MAC) operations. Other embodiments may be described and/or claimed.

## Claims

### Claim 1

A probabilistic in-memory computing apparatus for a stochastic neural network (SNN), the apparatus comprising:
a plurality of input lines arranged parallel to one another;
a plurality of accumulation lines arranged parallel to one another and intersecting the plurality of input lines; and
a plurality of probabilistic bit devices (p-bits), each p-bit of the plurality of p-bits is coupled to an input line of the plurality of input lines and an accumulation line of the plurality of accumulation lines, and
wherein each p-bit is a time-varying resistance device having a probabilistic resistance value based on a probability distribution of the SNN.

### Claim 2 (depends on 1)

The apparatus of claim 1, wherein the plurality of input lines are arranged perpendicular to the plurality of accumulation lines in a lateral or longitudinal plane.

### Claim 3 (depends on 2)

The apparatus of claim 2, wherein:
each input line of the plurality of input lines is laterally separated from other input lines of the plurality of input lines,
each accumulation line of the plurality of accumulation lines is laterally separated from other accumulation lines of the plurality of accumulation lines, and
each input line is longitudinally separated from each accumulation line.

### Claim 4 (depends on 2)

The apparatus of claim 2, wherein each p-bit includes a first terminal and a second terminal, wherein the first terminal of each p-bit is connected to a corresponding input line of the plurality of input lines and the second terminal of each p-bit is connected to a corresponding accumulation line of the plurality of accumulation lines.

### Claim 5 (depends on 4)

The apparatus of claim 4, wherein the second terminal of each p-bit is connected to a portion of the corresponding accumulation line that precedes an intersection point between the corresponding accumulation line and another input line of the plurality of input lines to which another p-bit of the plurality of p-bits is connected.

### Claim 6 (depends on 2)

The apparatus of claim 2, wherein:
a first input voltage to be applied to a first input line of the plurality of input lines is multiplied by a conductance of a first set of p-bits of the plurality of p-bits connected to the first input line, and is to cause first current to flow through respective accumulation lines to which the first set of p-bits are connected,
a second input voltage to be applied to a second input line of the plurality of input lines is multiplied by a conductance of a second set of p-bits of the plurality of p-bits connected to the second input line, and is to cause second current to flow through respective accumulation lines to which the second set of p-bits are connected, and
the first current is to accumulate with the second current at respective intersection points of the second input line and the respective accumulation lines.

### Claim 7 (depends on 1)

The apparatus of claim 1, wherein accumulation of current along each accumulation line of the plurality of accumulation lines is a respective multiply-and-accumulate (MAC) operation.

### Claim 8 (depends on 1)

The apparatus of claim 1, further comprising:
a plurality of analog-to-digital converters (ADCs), wherein each ADC of the plurality of ADCs is connected to a corresponding ADC of the plurality of accumulation lines, and each ADC is configured to convert accumulated current from the corresponding ADC into a respective digital signal.

### Claim 9 (depends on 1)

The apparatus of claim 1, wherein the probabilistic resistance value of individual p-bits of the plurality of p-bits is an element of a weight matrix of the SNN, and the weight matrix comprises weights for one layer of the SNN.

### Claim 10 (depends on 1)

The apparatus of claim 1, wherein the time-varying resistance device is a spin-orbit torque (SOT) tuned magnetic tunnel junction (MTJ) device comprising:
a spin-orbital coupling (SOC) layer;
a top barrier layer disposed on the SOC layer;
a free ferromagnetic layer disposed on the top barrier layer;
a tunneling barrier layer disposed on the free ferromagnetic layer; and
a reference ferromagnetic layer disposed on the tunneling barrier layer, and
wherein the free ferromagnetic layer is more magnetically isotropic than the reference ferromagnetic layer, or the reference ferromagnetic layer is more magnetically anisotropic than the free ferromagnetic layer.

### Claim 11 (depends on 10)

The apparatus of claim 10, wherein the SOT tuned MTJ device further comprises:
positive and negative bias terminals coupled to the SOC layer, wherein a current that flows between the positive and negative bias terminals is to bias the free ferromagnetic layer according to a mean of the probability distribution;
a first voltage input terminal coupled to the free ferromagnetic layer; and
a second voltage input terminal coupled to the reference ferromagnetic layer,
wherein a magnetoresistance is created between the positive voltage input terminal and the negative voltage input terminal based on the bias of the free ferromagnetic layer, and the probabilistic resistance value is based on the magnetoresistance.

### Claim 12 (depends on 1)

The apparatus of claim 1, wherein the time-varying resistance device is a ferroelectric gate tuned field effect transistor (FEFET) comprising:
a substrate including a source region, a drain region, and a channel region between the source region and the drain region;
a dielectric layer formed on the channel region of the substrate;
a ferroelectric (FE) layer formed on the dielectric layer; and
a top electrode (TE) layer formed on the FE layer.

### Claim 13 (depends on 12)

The apparatus of claim 12, wherein the FEFET further comprises:
a bias terminal coupled to the TE layer, wherein a bias signal to be applied to the bias terminal is to change an FE polarization of the FE layer;
a first voltage input terminal coupled to the source region; and
a second voltage input terminal coupled to the drain region,
wherein a channel resistance is created in the channel region based on the bias of the FE polarization, and the probabilistic resistance value is based on the channel resistance.

### Claim 14 (depends on 1)

The apparatus of claim 1, wherein the time-varying resistance device comprises random number generator (RNG) circuitry comprising:
a set of digital p-bits, wherein each digital p-bit in the set of p-bits is to generate respective digital bits;
p-bit interface circuitry connected to the set of digital p-bits, wherein the p-bit interface circuitry is to apply respective bias values to each digital p-bit of the set of digital p-bits;
a resistance bank connected to the set of digital p-bits and connected to one or more p-bits of the plurality of p-bits, wherein the resistance bank is configured to determine, based on the respective digital bits, the probabilistic resistance value of the one or more p-bits to which the resistance bank is connected.

### Claim 15 (depends on 1)

The apparatus of claim 1, wherein the stochastic neural network is one or more of a Bayesian network (BN), a Bayesian belief network (BBN), a Bayesian neural network (BNN), a deep BNN (DBNN), a dynamic BN (DBN), a probabilistic graphical model (PGM), a Boltzmann machine, a restricted Boltzmann machine (RBM), a Hopfield network, and a convolutional deep belief network (CDBN).

### Claim 16

A method of probabilistic in-memory computing of a stochastic neural network (SNN) by a probabilistic in-memory computing apparatus, the apparatus comprising a plurality of input lines, a plurality of accumulation lines intersecting the plurality of input lines; and a plurality of probabilistic bit devices (p-bits) each of which is coupled to an individual input line of the plurality of input lines and an individual accumulation line of the plurality of accumulation lines, wherein each p-bit is a time-varying resistance device having a probabilistic resistance value based on a probability distribution of the SNN, the method comprising:
applying respective input voltages to respective input lines of the plurality of input lines such that:
the respective input voltages are multiplied by conductances of respective sets of p-bits of the plurality of p-bits connected to the respective input lines, and is to cause current to flow through respective accumulation lines to which the respective sets of p-bits are connected, and
the current flowing through the respective accumulation lines is to accumulate with current at respective intersection points of the subsequent input lines of the plurality of input lines and the respective accumulation lines.

### Claim 17 (depends on 16)

The method of claim 16, wherein accumulation of current along each accumulation line of the plurality of accumulation lines is a respective multiply-and-accumulate (MAC) operation.

### Claim 18 (depends on 16)

The method of claim 16, wherein the probabilistic resistance value of individual p-bits of the plurality of p-bits is an element of a weight matrix of the SNN, and the weight matrix comprises weights for one layer of the SNN.

### Claim 19 (depends on 16)

The method of claim 16, wherein the time-varying resistance device is a spin-orbit torque (SOT) tuned magnetic tunnel junction (MTJ) device comprising a spin-orbital coupling (SOC) layer; a top barrier layer disposed on the SOC layer; a free ferromagnetic layer disposed on the top barrier layer; a tunneling barrier layer disposed on the free ferromagnetic layer; a reference ferromagnetic layer disposed on the tunneling barrier layer, wherein the free ferromagnetic layer is more magnetically isotropic than the reference ferromagnetic layer, or the reference ferromagnetic layer is more magnetically anisotropic than the free ferromagnetic layer; positive and negative bias terminals coupled to the SOC layer; a first voltage input terminal coupled to the free ferromagnetic layer; and a second voltage input terminal coupled to the reference ferromagnetic layer, and wherein the method further comprises:
applying a current to the positive or the negative bias terminals, wherein:
the applied current flows between the positive and negative bias terminals to bias the free ferromagnetic layer according to a mean of the probability distribution, and
a magnetoresistance is created between the positive voltage input terminal and the negative voltage input terminal based on the bias of the free ferromagnetic layer, and the probabilistic resistance value is based on the magnetoresistance.

### Claim 20 (depends on 16)

The method of claim 16, wherein the time-varying resistance device is a ferroelectric gate tuned field effect transistor (FEFET) comprising a substrate including a source region, a drain region, and a channel region between the source region and the drain region; a dielectric layer formed on the channel region of the substrate; a ferroelectric (FE) layer formed on the dielectric layer; a top electrode (TE) layer formed on the FE layer; a bias terminal coupled to the TE layer; a first voltage input terminal coupled to the source region; and a second voltage input terminal coupled to the drain region, and wherein the method further comprises:
applying a bias signal to the bias terminal to change an FE polarization of the FE layer, wherein a channel resistance is created in the channel region based on the bias of the FE polarization, and the probabilistic resistance value is based on the channel resistance.

### Claim 21 (depends on 16)

The method of claim 16, wherein the time-varying resistance device comprises random number generator (RNG) circuitry comprising a set of digital p-bits, p-bit interface circuitry connected to the set of digital p-bits, and a resistance bank connected to the set of digital p-bits and connected to one or more p-bits of the plurality of p-bits, and wherein the method further comprises:
operating the p-bit interface circuitry is to apply respective bias values to each digital p-bit of the set of digital p-bits;
operating each digital p-bit in the set of p-bits to generate respective digital bits based on the respective bias values; and
operating the resistance bank to determine, based on the respective digital bits, the probabilistic resistance value of the one or more p-bits to which the resistance bank is connected.

### Claim 22

A probabilistic in-memory computing system, comprising:
a plurality of p-bit arrays, wherein each p-bit array of the plurality of p-bit arrays is connected to at least one other p-bit array of the plurality of p-bit arrays, and each p-bit array comprises:
a plurality of input lines configured to receive respective input voltages,
a plurality of accumulation lines intersecting the plurality of input lines, and
a plurality of probabilistic bit devices (p-bits), each p-bit of the plurality of p-bits is coupled to an input line of the plurality of input lines and an accumulation line of the plurality of accumulation lines, and wherein each p-bit is a time-varying resistance device having a probabilistic resistance value based on a probability distribution of the SNN; and
processor circuitry communicatively coupled to the plurality of p-bit arrays to provide data for modulation of the respective input voltages,
wherein the respective input voltages to be applied to the individual input lines of the plurality of input lines is to cause current to flow through p-bits of the plurality of p-bits connected to the individual input lines and is accumulated on corresponding accumulation lines of the plurality of accumulation lines.

### Claim 23 (depends on 22)

The system of claim 22, wherein the plurality of p-bits of each p-bit array is a same type of p-bit as each other p-bit of the plurality of p-bits.

### Claim 24 (depends on 22)

The system of claim 22, wherein at least one p-bit in at least one p-bit array of the plurality of p-bit arrays is a different type of p-bit than other p-bits in the at least one p-bit array.

### Claim 25 (depends on 22)

The system of claim 22, wherein p-bits in at least one p-bit array of the plurality of p-bit arrays are different type of p-bits than p-bits in at least one other p-bit array of the plurality of p-bit arrays.

### Claim 26 (depends on 22)

The system of claim 22, wherein the plurality of p-bits includes one or more of spin-orbit torque (SOT) tuned magnetic tunnel junction (MTJ) devices, ferroelectric gate tuned field effect transistors (FEFETs), and random number generator (RNG) circuitry.
