Chipmakers Model AI For Radio Access Networks

The chips that power and connect smartphones are now foundational to a disparate portfolio of daily tasks we take for granted, from accessing the internet to snapping a photo or asking Siri or Google if rain is in the forecast. Most people don’t think twice about the conflicting demands these tasks can place on semiconductors, but for engineers at leading chip manufacturers, this balancing act is a central part of the design process.

This job has become more challenging with the proliferation of artificial intelligence (AI) engines, and applications that leverage them. AI places demands on smartphone processors as well as the modems and processors in base station radios, which engineers are now training to transmit and receive as efficiently as possible in a given set of conditions.

“The quality of AI is to be able to handle large amounts of data and detect patterns and handle certain situations,” said Nico Kelling, vice president and head of the Center of Excellence Artificial Intelligence at Infineon. “The AI can make smart decisions that are situation-dependent. This principle applies also to connectivity sites.”

For smartphones, connectivity protocols are defined by the Third Generation Partnership Project (3GPP). Last year that group articulated a need to increase its focus on AI and ML, noting the requirement to jointly optimize various key performance metrics, which is becoming more challenging with 5G.

Use cases for AI in the RAN
With Release 17, 3GPP initiated a study on the applications of AI/ML to radio access networks. Called Enhancement of Data Collection, this study involved end user devices making reports to network equipment, said Juan Montojo, vice president of engineering at Qualcomm. The study project is turning into a work item for Release 18. Using the principles of self-organizing networks, but moving from a reactive to a proactive stance, algorithms train base stations to power down when they do not receive location updates from terminal devices, an indication that no one in the area needs the network.

Teaching base stations to power down when they are not needed has obvious implications for energy use management. In fact, energy management is the leading use case for AI in wireless networks, noted Balaji Raghothaman, chief technologist for network solutions at Keysight Technologies.

“Keysight considers energy savings and beam management to be the most promising areas for AI/ML to add significant value in the near-term,” said Raghothaman. “Some industry trials have already shown significant savings to be had from intelligent energy savings algorithms that turn off certain base stations or antennas in the network judiciously during non-peak hours. Similarly, the dynamic shaping of beams for optimizing coverage and capacity is a promising area.”

Beamforming, which brings AI/ML applications to the air interface, is part of another AI-related project at 3GPP. While the Enhancement of Data Collection work item involved end-user devices making reports to intelligent base stations, the new project involves two-way communication between base stations and end-user devices, which will use AI to jointly solve non-linear problems.

Among the papers submitted to Release 18 working groups focused on the physical layer, AI/ML was the most popular subject, accounting for some 200 papers. As rapporteur for the AI/ML working group, Montojo is charged with synthesizing the information into a go-forward plan that is agreeable to all parties.

Montojo explained the group is studying three use cases for AI/ML in the air interface — beamforming, channel state information (CSI) and positioning. “Qualcomm is planning to start thinking of having dedicated hardware for this type of AI/ML,” said Montojo, adding that this would mean devoting more resources to AI in the CPUs and GPUs. “In my opinion, not just Qualcomm but everybody will be growing it,” he predicted.

Intel also is talking about the importance of AI in semiconductor solutions for 5G. Speaking recently at Informa’s Big 5G Event in Austin, Texas, Cristina Rodriguez, vice president in Intel’s Wireless Access Network Division, called AI and ML “game changers for 5G evolution and the future of every wireless network standard.” She forecast these technologies will support link adaptation, traffic steering, service-level agreements and energy savings.

According to Rodriguez, AI in radio access networks will give service providers more granular, real-time control of network operations. Machine learning will enable networks to automatically configure resources to meet workload requirements in the most efficient manner.

Rodriguez cited Capgemini’s use of Intel’s Xeon processors in a test called Project Marconi, which demonstrated a 15% improvement in spectral efficiency with real-time predictive analysis.

Intel claims to have the only line of server CPUs with built-in AI acceleration, and is positioning its Xeon processors as the bedrock infrastructure of virtualized radio access networks. As network operators consider RAN virtualization, AI is an important element of the value proposition. Rodriguez notes the ability of AI to enable network slicing, or the dynamic allocation of network resources to an enterprise customer, in order to guarantee that customer a consistent level of service. Network slicing is seen as a potential new revenue source for operators, enabled by 5G and AI.

AI also has important implications for open RAN, which refers to the opening of interfaces between RAN elements so that network operators can mix and match vendor equipment.

“There has already been work in forums like the O-RAN Alliance, regarding use cases such as Massive Multi-user MIMO Optimization, traffic steering, energy savings, etc,” said Keysight’s Raghothaman. “The testing of such optimization algorithms based on the RIC (RAN Intelligence Controller) framework, will require the use of large-scale network digital twins. For late 5G and 6G, more advanced AI-based methods are being proposed that directly impact the physical and MAC layers. Distributed / federated learning techniques, as well as AI-native modules are likely to be key advancements.”

For Keysight, this will create new challenges in testing and validation, Raghothaman said.  He believes the wireless industry will need to collectively decide whether synthetic/emulated training data is effective for training AI engines and whether the emulated modules normally used for test configurations should also include AI functionality.

AI for the air interface
The interface for AI in the air has the potential to significantly improve wireless network performance, as measured by speeds, latencies, reliability, and power use.

All three of the use cases 3GPP is studying for AI — beamforming, CSI and positioning — impact the ways smartphones and other terminal devices connect to cellular networks as they move. In many environments, a phone will have more than one possible connection point, creating a problem with multiple solutions.

“That is where AI/ML thrives the most,” explained Montojo. “It learns from all this training what to do … when you start losing one cell, how early should you move to another? Or when you see a beam of a certain strength, how likely is that by zooming in and refining that beam you will find a very good candidate for being your serving beam?”

The need to standardize approaches is driven by the number of companies and technologies involved. No matter how advanced Snapdragon’s AI acceleration is, Qualcomm can’t consistently apply AI to the air interface unless partners like Ericsson and Nokia are using the same approach on the network side.

“The intelligence will be sometimes in the network, sometimes in the user equipment, sometimes in both,” said Montojo. “All of these AI/ML techniques will try, out of a very noisy measurement set, to identify the first arrival path.”

Beamforming may be an ideal use case for AI, because it is currently accomplished through an iterative process that does not capture information and apply it to future, similar scenarios. Adaptive array smart antennas already can change the direction and width of a transmitted beam based on the environment, but with AI on both the transmit and receive end, the process of refining beams could be shortened or even eliminated. Antennas could be trained to predict gain and select the appropriate beam.

Channel state information (CSI) is another application of AI to the air interface. Transmitters and receivers both estimate CSI, such as attenuation, power decay and scattering, but do not yet typically perform predictive analysis or share information.

Positioning is 3GPP’s third area of focus as it studies applications of AI/ML to the air interface. AI can be used in conjunction with 5G sidelink, a protocol that enables end user devices to communicate directly with one another, bypassing base stations. When two devices do not have a line-of-sight connection, algorithms can predict the arrival path of an incoming transmission.

5G sidelink and AI have clear implications for autonomous vehicles. Already, self-driving cars communicate with one another directly, but they don’t typically use 5G for this. That could change in the future, especially if 5G-connected cars can use AI to “find” one another.

The opportunity to make 5G, and eventually 6G, foundational to increasingly autonomous vehicles is a pillar of Qualcomm’s business strategy. The company already supplies silicon to most of the top 20 automakers, and CEO Cristiano Amon has predicted Qualcomm could increase its semiconductor sales to the automotive industry 10X as advanced driver-assistance systems proliferate. So for Qualcomm, driving AI into the 5G air interface is about much more than smartphones.

AI and power use
Infineon’s Kelling estimates AI has a role to play in helping the world meet 80% of the United Nations’ Environmental, Social and Governance (ESG) goals. He said that as more and more devices of all kinds are equipped with AI and wireless interfaces, they will be able to connect to sensors, which will feed them actionable information about their environment. By “giving human senses to machines” and teaching them to conserve energy and reduce waste, AI will play a vital role in protecting the planet, he said.

For wireless networks standards bodies, AI can be a double-edged sword when it comes to sustainability. “If you start throwing in complexity because you can do these millions or trillions of operations per second, of course that is going to have an impact,” Montojo said. “Basically, you need to feed that brain to do all those computations and that is going to take power.”

He noted that some AI/ML initiatives will prove too power-intensive to be worth the efficiency gains, while others such as beamforming may be worth the energy.

Base stations can be the most power-hungry elements of the network, which is why 3GPP’s Release 17 study looked at ways to train them to power down automatically when their services are not needed. Release 18 will study ways to train smartphones to do the same.

Slicing the AI market
NVIDIA, Intel, and Qualcomm are all heavily invested in the AI market, and could have roles to play as this technology moves into radio access networks.

Qualcomm dominates AI in mobile devices, with a focus on natural language processing (through Hugging Face), image enhancement, and extended reality. The chipmaker also has AI solutions for the cloud, IoT and automotive markets.

NVIDIA’s AI expertise is grounded in graphics and gaming, but the company is also a major player in data center AI, and has set its sights on telco networks, as well. The chipmaker recently launched an AI-on-5G platform to help operators deploy AI applications along with connectivity at the network edge.

Intel’s AI solutions include Gaudi 2 for training, and Greco for inferencing. The company’s dominance in data centers, coupled with its early lead in the virtualization of radio access networks, make it likely to be a strong competitor for AI solutions in the RAN.

IBM, one of the first entrants into the market for AI chipsets, is also one of the first to be tapped by a mobile network operator for an AI solution. Dish Wireless, the first operator in the world to build a greenfield 5G open RAN cloud-based network, will use IBM’s AI-powered automation and network orchestration software.

Finally, Google is poised to play in this market, having recently launched ASICs that are purpose-built for AI. The search giant recently joined the ORAN Alliance, with the goal of lending its machine learning capabilities to operators who want to build intelligent networks in the cloud.

Creating consensus
The increasing application of AI and ML to smartphones, base stations, automobiles, and the communication links between all these is a development that ultimately will change the ways we use the machines we count on the most in our daily lives. A large number of stakeholders have to cooperate, and this will take time.

Fig. 1: Roadmap and features for different releases. Source: 3GPP

By the functional freeze of Release 18 in late 2023, 3GPP may have defined one or two applications for AI in the air interface, but actual implementation is not expected until Release 19. And the use cases 3GPP is studying may or may not be the ones that ultimately become part of the standard.

“The three use cases we selected are just pilots that we thought were meaningful and representative enough to expand to many other use cases down the road,” said Montojo. “If you fast forward a few years, everything will have an AI/ML component.”

Related
Increasing Performance With Data Acceleration
In-line acceleration boosts performance for radio access networks.
Open RAN Phase 2
Building a common, interoperable infrastructure for radio access networks.