Dynamically configurable, multi-ported co-processor for convolutional neural networks

• US 8,442,927 B2
• Filed: 02/01/2010
• Issued: 05/14/2013
• Est. Priority Date: 07/30/2009
• Status: Active Grant

• Current Assignee: NEC Corporation
• Original Assignee: NEC Laboratories America Inc (NEC Corporation)
• Inventors: Chakradhar, Srimat, Sankaradas, Murugan, Jakkula, Venkata S., Cadambi, Srihari
• Primary Examiner(s): Wong, Lut
• Application Number:
  US12/697,823
• Publication Number:
  US 20110029471A1
• Time in Patent Office: 1,198 Days
• Field of Search: 706/22, 706/42, 712/15, 712/34
• US Class Current: 706/22
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/063 using electronic means
• CPC Class Codes: G06N 3/084 Backpropagation, e.g. using gradient descent

Abstract

A coprocessor and method for processing convolutional neural networks includes a configurable input switch coupled to an input. A plurality of convolver elements are enabled in accordance with the input switch. An output switch is configured to receive outputs from the set of convolver elements to provide data to output branches. A controller is configured to provide control signals to the input switch and the output switch such that the set of convolver elements are rendered active and a number of output branches are selected for a given cycle in accordance with the control signals.

Claims

1. A method for processing convolutional neural networks (CNN), comprising:

determining a workload for a convolutional neural networks;

generating a control signal based upon the workload to permit a selection of a type of parallelism to be employed in processing a layer of the CNN;

configuring an input switch to enable a number of convolvers which convolve an input in accordance with the control signal;

configuring an output switch to enable a number of output branches for a given cycle in accordance with the control signal; and

processing outputs from the output branches, said processing including reconfiguring the input switch and the output switch in accordance with a next layer of the CNN to be processed.

2. The method as recited in claim 1, further comprising selecting the number of convolvers and the number of output branches based upon a type of parallelism needed for a given neural network layer.

3. The method as recited in claim 2, wherein the type of parallelism includes one of parallelism within a convolution operation, inter-output parallelism and intra-output parallelism.

4. The method as recited in claim 1, further comprising providing a memory subsystem having at least two banks, wherein a first bank provides input storage and a second bank provides output storage to enable a stateless, streaming coprocessor architecture.

5. The method as recited in claim 4, further comprising providing a third memory bank for storing intermediate results.