Search for: All records

With the rapid advance in Deep Neural Networks (DNNs), GPU’s role as a hardware accelerator becomes increasingly important. Due to the GPU’s significant power consumption, developing high- performance and power-efficient GPU systems is a critical challenge. DNN applications need to move a large amount of data between memory and the processing cores, which consumes a great amount of NoC power. However, prior proposed lossless data compressions cannot achieve optimal performance and energy efficiency because they did not take advantage of the error resilience of DNNs. In this work, we propose an NoC architecture that can reduce power consumption without compromising performance and accu- racy. Our technique takes advantage of the error resilience of DNNs as well as the data locality in the floating-point data representation of DNNs. Each data packet is reorganized by grouping data with similar bits such as in the exponents, and redundant bits are sent only once. We further compress the mantissa fields by appropri- ately selecting "proxy" values for data sharing the same exponent. Our evaluation results show that the proposed technique can ef- fectively reduce the amount of data transmitted and lead to better performance and power trade-offs while preserving accuracy.