NTT Docomo announced that it has successfully demonstrated running an AI application directly on general-purpose CPU resources of a virtual radio access network (vRAN) installed in a commercial network.
This demonstration confirmed the potential of network architecture designs to handle rapid traffic growth due to AI service expansion while optimizing network operating costs.
With the increasing adoption of AI-powered services such as generative AI and robotic control, network traffic is expected to increase dramatically, and data traffic is expected to explode in the coming years. Docomo is considering designing a next-generation network architecture called “in-network computing”*1 *2 that executes AI processing within the network. For telecommunications carriers who are entering the age of AI, it is important to continuously analyze various data within the network using AI to create new value, while at the same time improving energy efficiency and realizing sustainable network operations.
Docomo is conducting research on next-generation network architectures with the aim of maximizing the characteristics of each type of computing resource, including xPUs, to improve user experience, optimize network traffic, and achieve optimal placement and effective use of computing resources. Specifically, by appropriately placing high-performance GPUs, which are one of the key elements of advanced AI workloads, and CPUs suitable for wide-area deployment and efficient operation with low power consumption on the appropriate network nodes, we aim to improve performance and efficiency for customers from both the network and AI perspectives.
Docomo has integrated a platform that can simultaneously run vRAN functions and AI service applications by leveraging the CPU resources of general-purpose servers. As a result, it was confirmed that communication network processing and some AI processing can be executed in parallel even when using the CPU. This achievement demonstrates the flexibility of combining networks and AI applications without relying on dedicated high-performance accelerators, expanding the range of viable options for efficient network deployment.
Docomo will continue to consider and promote the optimal placement of computing resources such as CPUs and GPUs, taking into account actual traffic characteristics and the requirements of various AI applications.
This demo utilized a vRAN configuration consisting of the following commercial products:
vRAN base station software, provided by NEC Corporation
Amazon Web Services (AWS) provides virtualized infrastructure to host vRAN and AI applications.
An accelerator card provided by Qualcomm Technologies, Inc. that speeds up certain calculation processes.
Servers equipped with the above products provided by HPE
Docomo will continue to integrate and verify both network infrastructure and AI service platforms in order to realize “in-network computing,” and will continue to consider commercialization. In addition, Docomo will work with business partners to realize the optimal network architecture design for the 6G and AI era.
