Neural network processor using compression and decompression of activation data to reduce memory bandwidth utilization

摘要

The deep neural network (DNN) module may compress and decompress neurally generated activation data to reduce the utilization of memory bus bandwidth. The compression unit may receive an uncompressed chunk of data generated by neurons in the DNN module. The compression unit produces a mask portion and a data portion of the compressed output chunk. The mask portion encodes the presence and position of zero and nonzero bytes in the uncompressed chunk of data. The data portion stores truncated nonzero bytes from the uncompressed chunks of data. The decompression unit may receive a compressed chunk of data from the memory of the DNN processor or the memory of the application host. The decompression unit decompresses the compressed chunks of data using the mask portion and the data portion. This may reduce memory bus utilization, allow the DNN module to complete processing operations more quickly, and reduce power consumption.