Back to results
US 12,718,095 B2Granted
Processing apparatus for neural network calculation and method for operating the same
AssigneeDeep ET
InventorYong Beom Cho
Life of this patent
Estimating the term…
Abstract
A processing apparatus for a neural network calculation and an operating method thereof are proposed. The processing apparatus for a neural network calculation according to an exemplary embodiment includes a plurality of processor units which performs the learning of a neural network, a plurality of weight memories provided to store a weight corresponding to each of a plurality of layers which forms the neural network, and a main memory which acquires and stores the weight updated from the plurality of weight memories when the learning is completed.
Claims5
A processing apparatus for a neural network calculation, comprising:
- a plurality of processor units which performs learning of a neural network;
- a plurality of weight memories provided to store a weight corresponding to each of a plurality of layers which forms the neural network;
- a main memory which acquires and stores the weight updated from the plurality of weight memories when the learning is completed;
- and
- a controller which generates a first control signal associated with at least one of a calculation process for generating output data based on input data of any one layer of the plurality of layers and a target weight acquired from any one weight memory among the plurality of weight memories corresponding to the any one layer of the plurality of layers and an updating process for updating and storing the target weight in the any one weight memory by performing back propagation based on the output data and a second control signal associated with data transfer between the plurality of weight memories and the main memory,
- wherein each of the plurality of processor units includes:
- a calculation unit which performs the calculation process;
- and
- a weight updating unit which performs the updating process,
- wherein the controller generates the second control signal only after completing the learning to suppress a transition between the main memory and the plurality of weight memories before completing the learning,
- wherein the controller generates the first control signal including a mapping signal which allocates a mapping relationship between the plurality of processor units and the plurality of weight memories based on feature information of the neural network, and
- wherein the plurality of weight memories partitions a single memory, and a partition structure of the single memory is determined based on a number of layers and a number of neurons of the neural network.
9 more elements
A method for operating a processing apparatus for a neural network calculation, comprising:
- performing, by a plurality of processor units of the processing apparatus, learning of a neural network by acquiring a weight from a plurality of weight memories which stores the weight corresponding to each of a plurality of layers which forms the neural network, by a plurality of processor units;
- acquiring and storing the weight updated from the plurality of weight memories in a main memory when the learning is completed;
- generating, by a controller of the processing apparatus, a first control signal associated with at least one of a calculation process for generating output data based on input data of any one layer of the plurality of layers and a target weight acquired from any one weight memory among the plurality of weight memories corresponding to the any one layer and an updating process for updating and storing the target weight in the any one weight memory by performing back propagation based on the output data;
- and
- generating, by the controller, a second control signal associated with a data transfer between the plurality of weight memories and the main memory,
- wherein the performing the learning of the neural network includes:
- performing the calculation process;
- and
- performing the updating process,
- wherein the generating the second control signal is performed after completing the learning to suppress a transition between the main memory and the plurality of weight memories before completing the learning, and
- wherein the plurality of weight memories partitions a single memory, and a partition structure of the single memory is determined based on a number of layers and a number of neurons of the neural network.
8 more elements
Dependencies are inferred from the claim text.