As standards bodies like 3GPP and the O-RAN Alliance work to help operators integrate AI and machine learning into network architecture, 5G operators are working to understand how these technologies can improve network efficiency. It’s becoming more apparent that AI and machine learning can power automation, replacing tedious and error-prone manual tasks. And anything that can boost infrastructure utilization is bound to help an operator’s bottom line. But many questions remain, because AI is unexplored territory for most operators, many of whom haven’t started their first machine-learning project, let alone developed in-house expertise.
Critical Design Questions when Planning a 5G AI Project
As with most significant projects, a little planning up front when applying AI to the 5G network goes a long ways toward ensuring success at the end. The following list describes some of the design considerations operators can explore as they begin or extend their journey to utilizing AI in the network.
• What do we use as inputs? Especially for 5G networks, which use network slicing to deliver different quality of service to different sets of users or types of traffic, inputs can be quite varied. Example inputs include network measurements, resource usage, traffic loads and types and trouble indicators.
- What should we expect as outputs? The nature of outputs depends on what part of the network AI is being applied to, and what use case is being explored. Two popular use cases are network automation and resource optimization, which use different outputs to achieve different goals.
- Is performance important? To meet the typical five-nines reliability goals of most operators, AI and machine-learning models must provide the right answers, often in near real time. Therefore, the design of the AI system should include appropriate AI optimization technologies, such as high-performance CPUs, hardware accelerators, acceleration libraries, and optimized AI frameworks.
- Which machine-learning model should we use? There is a myriad of models available, all of which do different things best. Time series prediction models are popular, as are models that combine convolutional neural networks (CNNs) and recurrent neural networks (RNNs). Other choices include multilayer perceptron (MLP) models and reinforcement learning models. Because AI is a fast-changing field, operators may want to invest in a data scientist to keep up with new developments and help prevent costly mistakes.
- How do we go about training the AI system? A network AI system is generally trained with network data from the operator itself. Anonymized data can be shared with model vendors, for off-premises model training.
- What are implementation best practices? The implementation of a 5G AI project involves several steps, including model evaluation, hardware and software procurement, input and output integration, development of a change-management system and integration with the network’s physical compute, storage and network infrastructure. Working with a company such as Intel, which has strong relationships throughout the telecom industry, can help minimize risks and speed time to market.
Intel Can Help Operators with Their 5G AI Projects
Intel offers many technologies, libraries, toolkits and resources to maximize AI performance, whether it’s model training or actually running the model to generate real-world network optimizations. For example:
- Intel® Xeon® Scalable processors specialized for networking and network functions virtualization (NFV). These processors can improve machine-learning training and inference performance through the use of innovations such as Intel® Advanced Vector Extensions 512 and Intel® Deep Learning Boost.
- A complete set of complementary toolkits, called Intel® oneAPI, that simplify programming and help improve efficiency and innovations.
- An open-source unified data analytics and AI platform called Analytics Zoo. Developed by Intel, this platform simplifies scaling a range of AI models on a big data cluster like Apache Spark. It also provides versions of deep-learning frameworks that have been optimized for Intel® architecture.
- Hardware accelerators such as Intel® FPGA products, Intel® Movidius™ Vision Processing Units (VPUs) and eASIC solutions.
- Memory extension using Intel® Optane™ persistent memory (PMem), which can affordably expand system memory to avoid I/O bottlenecks.
Besides providing high-performance hardware and optimized software, Intel also works with operators and solution vendors to help enable 5G AI projects. For example, the Intel® Network Builders University provides operators with solution guidance and training, which can increase operators’ confidence as they develop AI systems. Operators can take advantage of Intel’s relationships across a wide range of Internet service providers (ISPs), telecommunications equipment manufacturers (TEMs) and platform vendors.
Ready to Learn More?
Designing and implementing an AI system is a complex task. From data collection to taking action on outputs, Intel technology and collaborative expertise can help operators fine-tune their AI project at every state. To learn more, read the white paper, “AI in the 5G Network: Six Questions You Weren’t Supposed to Ask.”