New Algorithm Compresses On-device AI Data
Samsung Electronics has unveiled a new technology that can greatly enhance the performance of the neural processing unit (NPU), one of the core items in the company’s vision to become the global No. 1 in the system semiconductor field by 2030. The new technology compresses the key data based on deep learning and increases artificial intelligence (AI) operation speed by about eight times, while greatly reducing power consumption.
Samsung Electronics announced on July 2 that the Samsung Advanced Institute of Technology recently showed the world-class on-device AI data compression algorithm at Computer Vision and Pattern Recognition (CVPR), the world's largest conference on computer vision.
The key of this technology is to reduce the size of deep learning data by one-eighth while maintaining accuracy when the semiconductor recognizes a specific situation. It aims to only quantize the significant range of the whole information through learning, thereby increasing the operation speed and reducing the power consumption at the same time.
The company’s experimental results show it can obtain the highest accuracy compared to other technologies by processing a 32-bit deep learning algorithm at 4 bits by selecting the core data. In addition, the same computation results were obtained when only 1/40 to 1/120 of the semiconductor transistors were used.
Samsung Electronics said, "Unlike ordinary AI devices that go through cloud servers to perform complex operations, this technology allows hundreds or thousands of operations to be performed on-device. The core of this technology is the high-speed computation of large volumes of data without consuming much power."
The company emphasized that this on-device AI technology reduces the cost of building a cloud server, while providing fast and stable operation through self-computation in autonomous driving and virtual reality (VR) situations at the same time. The company plans to increase its competitiveness by applying the on-device AI algorithm to mobile systems on chip (SoC) such as Exynos series as well as memory and sensors.
Meanwhile, Samsung Electronics said it will increase the size of its expert workforce by 10 times to 2,000 people and carry out necessary M&As in order to develop original technologies in commercializing a human brain-like AI semiconductor NPU.