Distributed computational system and method for artificial neural network
According to an example of the present disclosure, a neural processing unit (NPU) capable of encoding is provided. The NPU comprises one or more processing elements (PEs) which perform operations for a plurality of layers of an artificial neural network and generate a plurality of output feature maps. The NPU also comprises an encoder which encodes at least one particular output feature map among a plurality of output feature maps into a bitstream and then transmits thereof.
1. A neural processing unit (NPU) capable of encoding, the NPU comprising:
at least one processing element (PE) to perform operations for a plurality of layers of an artificial neural network and to generate a plurality of output feature maps; and
an encoder to encode at least one particular output feature map into a bitstream,
wherein the at least one particular output feature map is selectively encoded among the plurality of output feature maps,
wherein the at least one particular output feature map is selected based on at least one of a size and a kind of machine analysis task, and
wherein the at least one particular output feature map and reference information for reconstructing the at least one particular output feature map are transmitted via the bitstream.
2. The NPU of claim 1 ,
wherein the at least one particular output feature map is selected from the plurality of output feature maps based on particular information.
3. The NPU of claim 2 ,
wherein the particular information is received from a server or another NPU that receives the bitstream.
4. The NPU of claim 2 ,
wherein the particular information includes identification information of the at least one particular output feature map or identification information of a layer corresponding to the at least one particular output feature map.
5. The NPU of claim 2 ,
wherein the particular information is generated based on information on a size of the at least one particular output feature map, information on a machine task performed on another NPU receiving the bitstream, or information on an artificial neural network model to be processed on the another NPU receiving the bitstream.
6. The NPU of claim 1 ,
wherein the bitstream is transmitted to enable another NPU to perform calculations for subsequent layers of the artificial neural network.
7. The NPU of claim 1 ,
wherein the bitstream is divided into a header and a payload, and
wherein the header includes the reference information, and the payload includes the at least one particular output feature map.
8. A neural processing unit (NPU) capable of decoding, the NPU comprising:
a decoder to receive a bitstream and to reconstruct the bitstream into at least one feature map for decoding; and
at least one processing element (PE) to perform operations of an artificial neural network model,
wherein the bitstream includes the at least one feature map and reference information for reconstructing the at least one feature map,
wherein the at least one feature map for decoding relates to a particular output feature map of a particular layer among a plurality of layers of the artificial neural network model,
wherein the particular output feature map is selectively encoded among a plurality of output feature maps with respect to the particular layer,
wherein the particular output feature map is selected based on at least one of a size and a kind of machine analysis task, and
wherein the at least one PE utilizes the at least one feature map for decoding into an input feature map of a subsequent layer of the particular layer.
9. The NPU of claim 8 ,
wherein the NPU transmits particular information which allows the particular layer or the particular output feature map to be selected by another NPU transmitting the bitstream.
10. The NPU of claim 9 ,
wherein the particular information includes identification information of the particular output feature map or identification information of a layer corresponding to the particular output feature map.
11. The NPU of claim 9 ,
wherein the particular information is generated based on information on a size of the particular output feature map, information on a machine task to be performed on the NPU, or information on the artificial neural network model.
12. The NPU of claim 8 ,
wherein the bitstream is divided into a header and a payload, and
wherein the header includes the reference information, and the payload includes the at least one feature map.
13. A system comprising:
a first electric device including a first semiconductor for a first neural processing unit, the first semiconductor being configured to process a first layer section of an artificial neural network model, the artificial neural network model comprising a plurality of layers; and
a second electric device including a second semiconductor for a second neural processing unit, the second semiconductor being configured to process a second layer section of the artificial neural network model,
wherein the first electric device is configured to encode a particular output feature map of a last layer of the first layer section and transmit the particular output feature map and reference information for reconstructing the particular output feature map to the second electric device,
wherein the particular output feature map is selectively encoded among a plurality of output feature maps with respect to the last layer of the first layer section, and
wherein the particular output feature map is selected based on at least one of a size and a kind of machine analysis task.
14. The system of claim 13 ,
wherein a total weight size of layers in the first layer section is relatively smaller than a total weight size of layers in the second layer section.
15. The system of claim 13 ,
wherein the first neural processing unit is plural, and
wherein the second semiconductor is configured to sequentially process each of output feature maps outputted from the first neural processing unit.
16. The system of claim 13 ,
wherein the second semiconductor is configured to reuse a weight corresponding to at least a number of the first neural processing unit.
17. The system of claim 13 ,
wherein a size of a first memory of the first neural processing unit is relatively smaller than a size of a second memory of the second neural processing unit.
18. The system of claim 13 ,
wherein the first layer section and the second layer section of the artificial neural network model are determined based on a size of an output feature map of a particular layer.
19. The system of claim 13 ,
wherein the first layer section and the second layer section of the artificial neural network model are determined based on a size of a sum of weights of the first layer section and the second layer section, respectively.
20. The system of claim 13 ,
wherein a size of a first weight of the first layer section of the artificial neural network model is relatively smaller than a size of a second weight of the second layer section.