Keys to Building a Telecommunications Infra that Supports AI


Modern devices have vastly superior computing power when compared to their predecessors. The increased computing capacity comes with the additional benefit of low power consumption. 

New devices can process large volumes of data at a fraction of the cost required by previous generations of hardware.

Advanced technologies like the Internet of Things (IoT) and 5G networking allow telecommunications industries to scale their data processing capacities at an unprecedented rate.

AI in Telecom

Artificial intelligence (AI) is the field of computer science in which algorithms are used to automate computers to solve complex problems without human intervention.

With the help of artificial intelligence, telecom companies can automate routine tasks to increase operational efficiencies. Companies are now able to remove manual labor, thus minimizing human error. 

Client data processing costs can also be outsourced to an edge-computing network that prepares inputs for advanced decision-making algorithms in the main network. The combination of AI technologies with IoT, 5G, and edge computing enables companies to optimize network performance, energy efficiency, latency, and security requirements.

As AI is a data-driven technology, there is a need for telecom companies and their clients to focus on good data management practices to help keep operations smooth and in compliance with industry standards.

Data Centers and Distributed Networks: Key Components of AI Infrastructure in Telecom

For telecom networks, one of the biggest challenges in adopting AI is the construction of the network infrastructure. Earlier, telecom networks were only designed for telephony, but many networks have switched to modern 4G/LTE technologies that harness the power of digital signals for operational tasks. 

However, AI technologies require the allocation of additional computing resources to manage data inputs for training and inference processing. These resources are typically dependent upon computing servers placed in physical locations called data centers, and networking and application servers that transmit data and run front-end applications.

Hardware Servers vs. Virtual servers

Physical servers provide a large number of computing resources that can be completely dedicated to a single application or client resource. They are perfect for undertaking large data workloads and cases where data privacy is important. 

Yet, physical servers need physical storage that comes at a premium.

Virtualization is a way to run a computer entirely within an existing hard drive (also called a VM). A physical server can run multiple VMs that can be customized to client specifications. While VMs may not be as powerful as physical servers, they offer a lighter, scalable solution for running applications dependent on algorithms – like AI and machine learning (ML) applications.

Distributed Computing Resources

In real-world scenarios, computer systems in telecom networks have limited bandwidth and computing resources. This means they can only process and transmit a limited volume of data, and perform a limited number of calculations in a given time. 

Distributed networks attempt to offset these limits by allocating core tasks to a central server that has sufficient bandwidth and computing resources. This central server meets the low latency requirements and high processing speeds. It can be a cloud server.

Computers that are short of either bandwidth or computing power are separated from the central server. These devices can be placed closer to the main server if they have enough bandwidth, or near application servers if low latency is required. These computers form the edge network. The requirements for a specific configuration of the distributed network depend on the use case for an application, as we will demonstrate below.

Centralized Training and Inference

Centralized servers offer high computing resources and bandwidth in a location. They are usually placed far from the application layer. A centralized server will have AI/ML inputs to be relayed from another location. 

After the algorithm has processed the inputs and computed an inference, it relays them back to the application server. Centralized resources are required when: 

  • Large amounts of data need to be processed
  • The processing is complex in nature
  • Latency or real-time efficiency is not important

Distributed Training and Inference

Distributed computers for running training and inference models lie at the opposite ends of a network. The algorithms running on these computers need to be simplified to run on limited hardware resources. Data from these computers will be processed closer to the application and can optionally be sent to a central server for additional computing. Distributed systems like this require additional work to synchronize them with the main network. Distributed computing resources are required when:

  • Ultra-low latency is required for performance or user experience
  • Data inputs are simple
  • Computing resources or bandwidth is limited.

Hybrid (Centralized/Distributed) Training and Inference

In telecoms, infrastructure design can be designed with flexibility in mind for a wide variety of applications. Because telecom companies deal with big data (large volumes of structured and unstructured data), there is a need to split computing resources so that latency, performance, and security are not compromised. 

For example, a telecoms company that wants to have the lowest latency in its services to its customers, will have to compromise on performance and security. On a practical level, this is not possible, and the requirements for acceptable latency lie between acceptable low and high ranges.

This is where hybrid infrastructure is useful. Some AI/ML modeling techniques include:

Extract, Transform, Load

An extract, transform, load (ETL) model is used when edge devices are not very powerful and a central server is needed. In these cases, edge computers send data to a central server in batches for processing. 

The central server prunes the data, simplifying its structure, and sends it back to the edge computers for training the AI. This technique eases the computational burden on edge devices and allows them to train AI without requiring high computing resources at all times.

Centralized Initial Training

An exclusively centralized initial training that removes the burden on edge-computing networks. Less resource-intensive retraining can be transferred to the edge-network.

Reducing the data size for meeting bandwidth and computing requirements can be achieved in two ways:

  • Quantization. This is the process of lowering the precision of input values sent for algorithmic computation so that the data size is reduced.
  • Sparsification. This refers to minimizing the coding length of stochastic gradients in a neural network, reducing its memory and computation costs.

Balancing Pipeline Complexity and Maintenance Costs

An AI/ML pipeline typically has multiple stages of model experimentation, testing, and deployment that require manual work. Generally, pipeline complexity is proportional to the number of features in your AI model. 

To minimize costs for work performed by supporting operations staff (data scientists, engineers, and developers), some of the steps in the pipeline should be automated. Implementing an ML process that is not automated will significantly increase development and operational costs. Teams should consider reducing a feature set if an AI implementation is proving to be too costly.

Build Better Telecom with AI

5G and IoT systems enable telecoms to create distributed systems that allow for cost-effective and scalable AI solutions. Still, these benefits are only realized through careful and innovative implementations of hardware infrastructure that adequately supports AI algorithms. 

By understanding the full potential and limits of edge-computing and cloud networks, the telecoms sector can use smart infrastructure to create AI systems that add value to both businesses and consumers. 

About the author:

Subbu Seetharaman is the Director Of Engineering, at Lantronix, a global provider of turnkey solutions and engineering services for the internet of things (IoT). Subbu is an engineering executive with over 25 years experience in leading software development teams, building geographically distributed, high performing teams involved in developing complex software products around programmable hardware devices.