In recent years, the demand for energy-efficient computing has been steadily increasing. This is especially true for mobile devices, where energy efficiency is a major concern. As such, researchers have been looking for ways to reduce energy consumption while still providing high performance. One promising approach is the use of dynamic neural networks (DNNs) on heterogeneous multi-processor system-on-chip (MPSoC) architectures.
A DNN is a type of artificial neural network that is capable of adapting to changing input data. This makes them ideal for applications such as image recognition, natural language processing, and autonomous driving. However, DNNs require a large amount of computational power, which can be difficult to provide on mobile devices. To address this issue, researchers have developed energy-efficient execution schemes for DNNs on MPSoC architectures.
The goal of these schemes is to minimize energy consumption while still providing high performance. To achieve this, they employ various techniques such as task scheduling, data partitioning, and dynamic voltage and frequency scaling (DVFS). Task scheduling is used to assign tasks to processors in an optimal manner, while data partitioning splits the data among multiple processors to reduce communication overhead. Finally, DVFS dynamically adjusts the voltage and frequency of the processors to reduce energy consumption.
In addition to these techniques, researchers have also proposed various optimization techniques to further reduce energy consumption. These include model compression, which reduces the size of the model by removing redundant parameters; weight pruning, which removes unnecessary weights from the model; and quantization, which reduces the precision of the weights to reduce memory requirements.
Overall, energy-efficient execution schemes for DNNs on MPSoC architectures have been shown to be effective in reducing energy consumption while still providing high performance. By employing various techniques such as task scheduling, data partitioning, and DVFS, as well as optimization techniques such as model compression, weight pruning, and quantization, researchers have been able to significantly reduce energy consumption while still providing high performance. As such, these schemes are becoming increasingly popular for mobile applications where energy efficiency is a major concern.
Source: Plato Data Intelligence: PlatoAiStream