Researchers from various institutions, including The Hong Kong University of Science and Technology, Fudan University, and the University of Electronic Science and Technology of China, have developed a new hardware accelerator designed to improve the energy efficiency of artificial intelligence (AI) tasks crucial for the energy sector, such as semantic segmentation. This technology could enhance applications like renewable energy infrastructure monitoring and smart grid management.
The team presented a CNN-Transformer accelerator fabricated using 28nm technology, occupying an area of 13.93 square millimeters. This accelerator is designed to perform semantic segmentation tasks, which involve classifying each pixel in an image to enable machines to understand and interpret visual data. The researchers claim that their design achieves a significant energy reduction of 3.86 to 10.91 times compared to previous designs. This improvement is attributed to several innovative features, including a hybrid attention unit, a layer-fusion scheduler, and a cascaded feature-map pruner. The peak energy efficiency of the accelerator is reported to be 52.90 tera operations per second per watt (TOPS/W) using 8-bit integer (INT8) operations.
The hybrid attention unit allows the accelerator to focus on relevant parts of the input data, improving the efficiency of the convolutional neural network (CNN) and Transformer models. The layer-fusion scheduler optimizes the sequence in which different layers of the neural network are processed, reducing redundant computations. The cascaded feature-map pruner eliminates unnecessary data early in the processing pipeline, further saving energy.
For the energy sector, this technology could be particularly useful in applications requiring real-time image processing and analysis, such as monitoring solar panels for defects or inspecting wind turbines for damage. By reducing the energy consumption of AI tasks, this accelerator could help make renewable energy infrastructure more efficient and cost-effective. The research was published in the IEEE Journal of Solid-State Circuits, a reputable source for advancements in semiconductor and integrated circuit design.
This article is based on research available at arXiv.