US 12,443,835 B2
Hardware architecture for processing data in sparse neural network
Kevin Lee Hunter, Sunnyvale, CA (US); and Subutai Ahmad, Palo Alto, CA (US)
Assigned to Numenta, Inc., Redwood City, CA (US)
Filed by Numenta, Inc., Redwood City, CA (US)
Filed on May 27, 2021, as Appl. No. 17/332,181.
Claims priority of provisional application 63/087,641, filed on Oct. 5, 2020.
Prior Publication US 2022/0108156 A1, Apr. 7, 2022
Int. Cl. G06N 3/08 (2023.01); G06N 3/048 (2023.01); G06N 3/063 (2023.01)
CPC G06N 3/063 (2013.01) [G06N 3/048 (2023.01); G06N 3/08 (2013.01)] 16 Claims
OG exemplary drawing
 
1. An artificial intelligence accelerator for performing operations related to a sparse neural network, comprising:
a memory circuit configured to store a sparse weight tensor and an activation tensor corresponding to a node of the sparse neural network;
a sparsity processing circuit coupled to the memory circuit and comprising lanes, the sparsity processing circuit configured to:
divide the sparse weight tensor into a plurality of blocks or partitions, a total number of the plurality of blocks or partitions in the sparse weight tensor equal to a total number of the lanes, each of the lanes assigned to one of the plurality of blocks or partitions, and
each of the blocks or partitions comprising a same fixed number of active values,
determine one or more locations of one or more active values in the plurality of blocks or partitions, and
send the one or more active values in the plurality of blocks or partitions to the lanes without sending inactive values in the plurality of blocks or partitions to the lanes; and
a set of multiply circuits coupled to the lanes of the sparsity processing circuit,
each of the multiply circuits configured to receive, via the lanes, the one or more active values fetched based on the one or more locations of the one or more active values determined by the sparsity processing circuit and perform a mathematical operation between the one or more active values in the plurality of blocks or partitions and one or more values of the activation tensor.