# US 2022/0277190 A1 · Neural network engine with associated memory array

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

- Number: US 2022/0277190 A1
- Document ID: US20220277190A1
- Status: Published application
- Kind: A1
- Application number: 17/677,772
- Filed: 2022-02-22
- Published: 2022-09-01
- Earliest claimed: 2021-02-28
- Assignees: Anaflash Inc.
- Applicants: Anaflash Inc.
- Inventors: Seung-Hwan Song; Sihwan Kim
- Web page: https://patentgrep.com/patent/US20220277190A1

## CPC

- G06F 7/53: in parallel-parallel fashion, i.e. both operands being entered in parallel (G06F7/533 takes precedence)
- G06N 3/04: Architecture, e.g. interconnection topology
- G06N 3/0495: Quantised networks; Sparse networks; Compressed networks
- G06N 3/063: using electronic means
- G06N 3/065: Analogue means
- G06F 7/5443: {Sum of products (for applications thereof, see the relevant places, e.g. G06F17/10, H03H17/00)}
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

A neural network engine configured to perform neural network computations. The neural network engine includes a first memory block configured to provide a first operand; a second memory block configured to provide a second operand; a synapse array configured to produce partial products of the first and second operands, wherein the synapse array is located outside of the first and second memory blocks; and, an accumulator array to perform summation of the partial products produced by the synapse array.

## Claims

### Claim 1

A neural network engine to perform computations, comprising:
a first memory block configured to provide a first operand;
a second memory block configured to provide a second operand;
a synapse array configured to produce partial products of the first and second operands, wherein the synapse array is located outside of the first and second memory blocks; and,
an accumulator array to perform summation of the partial products produced by the synapse array.

### Claim 2 (depends on 1)

The neural network engine of claim 1 wherein the synapse array is connected to a first group of multiple parallel bus lines, wherein each of the parallel bus lines includes multiple parallel channels for parallel transmission of multiple data bits of the first operand from the first memory block to the synapse array, respectively.

### Claim 3 (depends on 2)

The neural network engine of claim 2 wherein the synapse array is connected to a second group of multiple parallel bus lines, wherein each of the parallel bus lines includes multiple channels for transmitting multiple bits of the second operand from the second memory block to the synapse array, respectively.

### Claim 4 (depends on 3)

The neural network engine of claim 3 wherein the synapse array comprises multiple parallel synapse circuits for performing multiple parallel multiplication operations of all bits of the first and second operands.

### Claim 5 (depends on 4)

The neural network engine of claim 4, wherein each of the synapse circuits is connected to the first and second parallel bus lines to simultaneously receive the bits of the first and second operands from the first and second memory blocks, respectively.

### Claim 6 (depends on 5)

The neural network engine of claim 5 wherein each of the synapse circuits comprises multiple parallel synapses, wherein each synapse is configured to simultaneously produce multiple partial products based on the data bits of the first and second operands.

### Claim 7 (depends on 6)

The neural network engine of claim 6 wherein each synapse circuit is connected to two groups of parallel channels such that a first group channels enables transmission of two complementary inputs of the first operand from the first memory block and a second group channels enables parallel transmission of multiple bits of the second operand from the second memory block.

### Claim 8 (depends on 7)

The neural network engine of claim 7 wherein the synapse circuit further comprises a sign selection unit adjacent to the synapse producing the most significant bit of the partial product, the sign selection unit comprising:
a pair of parallel input channels Px and Nx connected to the respective first group channels;
a pair of parallel output lines P and N connected to a plurality of the multiple parallel synapses arranged in a row of the synapse array;
an input port connected to the corresponding one of the second group channels for receiving one sign input bit of the second operand from the second memory block; and,
multiple switches operable in response to the sign input bit value of the second operand.

### Claim 9 (depends on 8)

The neural network engine of claim 8 wherein a first group of the multiple switches are configured to connect the input line Px to the output line N and the input line Nx to the output line P when the input port receives a sign bit value of 1 of the second operand.

### Claim 10 (depends on 9)

The neural network engine of claim 9 wherein a second group of the multiple switches are configured to connect the input line Px to the output line P and the input line Nx to the output line N when the input port receives a sign bit value of 0 from the second operand.

### Claim 11 (depends on 10)

The neural network engine of claim 10 wherein bit values on the (Px, Nx) channels consist of complementary values (1,0) or (0,1), or all zeros (0,0).

### Claim 12 (depends on 11)

The neural network engine of claim 11 wherein the synapse comprises four MOS translators, wherein each transistor comprises four MOS transistors to perform a 1-bit multiplication operation of the first and second operands.

### Claim 13 (depends on 12)

The neural network engine of claim 12 wherein first and second of the four MOS transistors are serially connected and have gates connected to the respective output lines P and N and connected to a pair of primary and complementary bit lines, respectively.

### Claim 14 (depends on 13)

The neural network engine of claim 13 wherein one end of the first transistor is coupled to a primary bit line (BL) and one end of the second transistor is coupled to a complementary bit line (BLB).

### Claim 15 (depends on 14)

The neural network engine of claim 14 wherein the first transistor is connected to the primary bit line BL to output a respective 1-bit partial product of the multiplication when the operands A and B have the same sign value; and wherein the second transistor is connected to the complementary bit (BLB) to output a respective 1-bit partial product of the multiplication when the operands A and B have a different sign value.

### Claim 16 (depends on 15)

The neural network engine of claim 15 wherein third and fourth of the four MOS transistors are serially connected, and wherein the third transistor has a gate terminal connected to one of the second group channels connecting the second memory block and the fourth transistor has a gate connected to a reference signal source Vref that can be an analog voltage source having a range of VDD and VSS.

### Claim 17 (depends on 11)

The neural network engine of claim 11 wherein the synapse comprising:
a first AND gate having a first input connected to the P output line, a second input connected to one of the second group channels, and an output connected to a primary of one paired bit lines; and,
a second AND gate having a first input connected to the N output line, a second input connected to the channel the first AND gate is connected, and an output connected to a complementary of the one paired bit lines.

### Claim 18 (depends on 1)

The neural network engine of claim 1, further comprising:
a synapse driver configured for connecting the internal memory to the synapse array, wherein the synapse driver is configured to convert a streaming raw input data into two or more pairs of input data to perform multiple 1-bit multiplication in parallel.

### Claim 19 (depends on 11)

The neural network engine of claim 11, wherein the synapse driver is configured to generate a pair of bits corresponding to 1-bit of the raw data response to a sign bit value of the raw data received.
