Fujitsu Laboratories Ltd. is utilizing local 5G (5th generation mobile communication system for specific areas) to transform the business of manufacturing sites, and has acquired a large amount of video data acquired by a large number of cameras installed throughout the vast factory. We have developed an automatic design technology for a system that can analyze data at high speed.
Effect of reducing system cost by linking edge and data center (Source: Fujitsu Laboratories Ltd.) With the technology developed this time, the processing performed by the GPU (Graphics Processing Unit) server is divided into containers by process, and it is an inexpensive edge server. And the GPU server integrated in the data center are efficiently linked and executed. Furthermore, by making it possible to set resource requirements such as CPU (Central Processing Unit) clock frequency and GPU capacity according to the amount of video analysis processing required for each container, it can be used for edge servers and data centers while maintaining high speed. It can be designed to be deployed automatically. As a result, the data center can absorb fluctuations in the video processing load on the edge servers at each site, and the cost of the entire system can be reduced by up to one-third without having to maintain the number of servers assuming peak hours. In the past, video analysis using artificial intelligence (AI) required installing an expensive GPU server at each site on a scale that would meet the peak processing load. Therefore, there is a problem that a huge equipment cost is required to analyze a large number of camera images installed in the entire factory. In addition, when the processing load of image analysis fluctuates greatly after the start of system operation, there are problems in terms of operation such as the system configuration cannot be easily changed, which makes it difficult to carry out large-scale image analysis. Regarding resource requirements, not only the required number of CPU cores and memory amount, but also CPU clock frequency, GPU performance, and other requirements specific to video processing that cannot be met by conventional automatic container deployment technology can be handled as parameters. On top of that, the deployment destination of each container is automatically determined from the servers of various specifications in the edge and the data center so that the traffic between the edge and the data center is the least and the resource requirements of each process are satisfied. To do. In addition, multiple containers can be executed simultaneously on one GPU, and can be scheduled according to the degree of real-time demand for processing in each container.
Design technology that optimizes the performance of a system that links the edge and the data center (Source: Fujitsu Laboratories Ltd.) Video of worker behavior taken with 16 Full-HD cameras in an experimental environment assuming an assembly plant Therefore, when we built a video analysis system that detects assembly work mistakes and stagnation of transported goods with AI, we designed a system that links 16 edge servers and a data center with this technology, and made work mistakes in a few seconds. It was confirmed that the cost of the entire system including the server can be reduced by up to one-third.