Vector computation unit in a neural network processor

Patent metadata

Field	Value
Publication number	US-12277499-B2
Application number	US-202418636640-A
Country	US
Kind code	B2
Filing date	Apr 16, 2024
Priority date	May 21, 2015
Publication date	Apr 15, 2025
Grant date	Apr 15, 2025

Abstract

A circuit for performing neural network computations for a neural network comprising a plurality of layers, the circuit comprising: activation circuitry configured to receive a vector of accumulated values and configured to apply a function to each accumulated value to generate a vector of activation values; and normalization circuitry coupled to the activation circuitry and configured to generate a respective normalized value from each activation value.

First claim

The invention claimed is:

1. A vector computation unit for performing neural network computations comprising:
normalization circuitry configured to:
receive a vector of activation values;
receive one or more control signals specifying a normalization function; and
apply the normalization function to the activation values to generate respective normalized values for each activation value.

2. The vector computation unit of claim 1 , wherein the one or more control signals are provided by a sequencer.

3. The vector computation unit of claim 1 , further comprising activation circuitry configured to:
receive a vector of accumulated values;
receive one or more control signal specifying an activation function; and
apply the activation function to the accumulated values to generate the vector of activation values.

4. The vector computation unit of claim 3 , wherein the accumulated values correspond to products of a matrix multiplication between a layer of the neural network and a parameter matrix for the layer.

5. The vector computation unit of claim 1 , further comprising pooling circuitry configured to:
receive the normalized values;
receive one or more control signals specifying a pooling function; and
apply the pooling function to the normalized values to generate a pooled value.

6. The vector computation unit of claim 5 , wherein the pooled value comprises at least one of a maximum, a minimum, or an average of the activation values, or a maximum, a minimum, or an average of a subset of the activate values.

7. The vector computation unit of claim 5 , wherein the pooling circuitry comprises multiple parallel pooling circuitries, each pool circuitry configured to receive a subset of the activation values to generate a respective pooled value.

8. The vector computation unit of claim 1 , wherein applying the normalization function to the activation values comprises:
determining a sum of the activation values;
determining a multiplication factor based on the sum of the activation values; and
multiplying the activation values by the multiplication factor to generate the normalized values.

9. The vector computation unit of claim 1 , further comprising a plurality of registers and a plurality of memory units configured to store the activation values.

10. A method for performing neural network computations comprising:
receiving, by normalization circuitry of a vector computation unit, a vector of activation values;
receiving, by the normalization circuitry, one or more control signals specifying a normalization function; and
applying, by the normalization circuitry, the normalization function to the activation values to generate respective normalized values for each activation value.

11. The method of claim 10 , wherein the one or more control signals are provided by a sequencer.

12. The method of claim 10 , further comprising:
receiving, by activation circuitry of the vector computation unit, a vector of accumulated values;
receiving, by the activation circuitry, one or more control signal specifying an activation function; and
applying, by the activation circuitry, the activation function to the accumulated values to generate the vector of activation values.

13. The method of claim 12 , wherein the accumulated values correspond to products of a matrix multiplication between a layer of the neural network and a parameter matrix for the layer.

14. The method of claim 10 , further comprising:
receiving, by pooling circuitry of the vector computation unit, the normalized values;
receiving, by the pooling circuitry, one or more control signals specifying a pooling function; and
applying, by the pooling circuitry, the pooling function to the normalized values to generate a pooled value.

15. The method of claim 14 , wherein the pooled value comprises at least one of a maximum, a minimum, or an average of the activation values, or a maximum, a minimum, or an average of a subset of the activate values.

16. The method of claim 14 , wherein the pooling circuitry comprises multiple parallel pooling circuitries, each pool circuitry configured to receive a subset of the activation values to generate a respective pooled value.

17. The method of claim 10 , wherein applying the normalization function to the activation values comprises:
determining a sum of the activation values;
determining a multiplication factor based on the sum of the activation values; and
multiplying the activation values by the multiplication factor to generate the normalized values.

18. A non-transitory computer readable medium for storing instructions executable by a processor to perform neural network computations, the instructions comprising:
receiving a vector of activation values;
receiving one or more control signals specifying a normalization function; and
applying the normalization function to the activation values to generate respective normalized values for each activation value.

19. The non-transitory computer readable medium of claim 18 , wherein the instructions further comprise:
receiving a vector of accumulated values;
receiving one or more control signal specifying an activation function; and
applying the activation function to the accumulated values to generate the vector of activation values.

20. The non-transitory computer readable medium of claim 18 , wherein the instructions further comprise:
receiving the normalized values;
receiving one or more control signals specifying a pooling function; and
applying the pooling function to the normalized values to generate a pooled value.

Assignees

• Google Llc

Inventors

• Thorson Gregory Michael
• Clark Christopher Aaron
• Luu Dan

Classifications

• G06N3/08 Primary

  Learning methods · CPC title

• G06N3/0499

  Feedforward networks · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N5/04

  Inference or reasoning models · CPC title

• G06F7/544

  for evaluating functions by calculation {(G06F7/4824 takes precedence)} · CPC title