# US 2023/0065730 A1 · Power management in processing circuitry which implements a neural network

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

- Number: US 2023/0065730 A1
- Document ID: US20230065730A1
- Status: Published application
- Kind: A1
- Application number: 17/863,094
- Filed: 2022-07-12
- Published: 2023-03-02
- Earliest claimed: 2021-08-17
- Assignees: Axis AB
- Applicants: Axis AB
- Inventors: Anton JAKOBSSON
- Web page: https://patentgrep.com/patent/US20230065730A1

## CPC

- G06F 1/324: by lowering clock frequency
- G06F 1/206: {comprising thermal management}
- G06F 1/3296: by lowering the supply or operating voltage
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/063: using electronic means
- Y02D 10/00: Energy efficient computing, e.g. low power processors, power management or thermal management
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- Y02D: CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE

## Abstract

A method of operating a hardware accelerator comprises: implementing a multi-layer neural network using the hardware accelerator; measuring a power consumption of the hardware accelerator while executing a predefined operation on the multi-layer network at a default clock frequency; evaluating one or more power management criteria for the measured power consumption; and, in response to exceeding one of the power management criteria, deciding to reduce the clock frequency relative to the default clock frequency. In the step of measuring a power consumption of the hardware accelerator, per-layer measurements which each relate to fewer than all layers of the neural network may be captured.

## Claims

### Claim 1

A method of operating a hardware accelerator, the method comprising:
implementing a multi-layer neural network using the hardware accelerator;
measuring a power consumption of the hardware accelerator while executing a predefined operation on the multi-layer neural network at a default clock frequency;
evaluating one or more power management criteria for the measured power consumption; and,
in response to exceeding one of the power management criteria, reducing the clock frequency relative to the default clock frequency.

### Claim 2 (depends on 1)

The method of claim 1, wherein the measuring includes performing per-layer measurements which each relate to fewer than all layers of the neural network.

### Claim 3 (depends on 1)

The method of claim 1, wherein the power management criteria include at least one of:
a global power budget on momentary power,
a global power budget on average power,
a per-layer momentary power limit,
a per-layer average power limit, and
a temperature limit.

### Claim 4 (depends on 1)

The method of claim 1, further comprising:
measuring the power consumption while executing the predefined operation at an alternative, lower clock frequency.

### Claim 5 (depends on 1)

The method of claim 1, further comprising:
recording a performance indicator during the execution of the predefined operation for two or more drive-voltage values and at the same clock frequency,
wherein the evaluation includes comparing the drive-voltage values with respect to the recorded performance indicator.

### Claim 6 (depends on 5)

The method of claim 5, wherein the performance indicator is one of:
an execution time in seconds, and
a throughput per unit power.

### Claim 7 (depends on 1)

The method of claim 1, further comprising:
executing a production-related operation on the multi-layer neural network using the default clock frequency or, if applicable, the reduced clock frequency.

### Claim 8 (depends on 7)

The method of claim 7, wherein said decision is made on the basis of a conversion factor relating the predefined operation and the production-related operation.

### Claim 9 (depends on 8)

The method of claim 8, further comprising retrieving the conversion factor from a memory co-localized with the hardware accelerator.

### Claim 10 (depends on 1)

The method of claim 1, wherein the predefined operation executing during the measuring is not production-related.

### Claim 11 (depends on 1)

The method of claim 1, wherein the predefined operation executing during the measuring is production-related.

### Claim 12 (depends on 7)

The method of claim 7, wherein the production-related operation is related to video surveillance.

### Claim 13 (depends on 1)

The method of claim 1, wherein the hardware accelerator is specified for continuous operation at the default clock frequency, such as the greatest frequency at which the hardware accelerator is specified for continuous operation.

### Claim 14 (depends on 1)

The method of claim 1, wherein the neural network is a convolutional neural network, CNN.

### Claim 15

A processor configured to control a hardware accelerator, comprising:
a memory; and
processing circuitry configured to perform a method of operating a hardware accelerator, the method comprising:
implementing a multi-layer neural network using the hardware accelerator;
measuring a power consumption of the hardware accelerator while executing a predefined operation on the multi-layer neural network at a default clock frequency;
evaluating one or more power management criteria for the measured power consumption; and,
in response to exceeding one of the power management criteria, reducing the clock frequency relative to the default clock frequency.
