Unleashing

the potential of
AI in the RAN
 Table of contents

Bringing intelligence to the RAN / 03

How AI and machine learning are used in RANs / 05

Understanding the intelligent RAN architecture / 07

Table of contents Demonstrations of AI in the RAN / 11

Saving energy through dynamic power management / 12

Improving spectrum efficiency / 14

Improving MU-MIMO efficiency / 16

Improving resource utilization of O-DU upper PHY software / 18

Improving traffic steering / 20

Accelerating AI on general-purpose processors / 23

Conclusion / 25

Empower Customers and Streamline Operations with Computer Vision at Checkout 2

Bringing intelligence to
the RAN
Mobile data consumption is increasing. For example, GSMA
predicts that mobile data traffic in Europe will almost triple
between 2023 and 20281, while CTIA reports that 2022 saw
the greatest increase in mobile data traffic ever in the US.2
It increased from 53.4T MB to 73.7T MB, nearly double the
previous year’s increase.

The industry introduced 5G in response to escalating data

demand and to enable new use cases. With this new standard
came more frequency bands, more radio types, more base
stations, and a wider variety of connected devices, resulting
in more network complexity. In addition, vendor complexity
is growing as the network becomes more open. Network
operators need novel ways to address these challenges.

Bringing intelligence to the RAN 3

 Artificial intelligence (AI) will play a major role in new telco
use cases and in optimizing 5G networks. Some operators are
already using it to optimize cell site design. The next frontier for
AI in the network is optimizing resources and user experience
in real-time using data about network conditions and user
behavior. In this way, AI can help mitigate network complexity,
reduce operating costs, and improve Quality of Experience
(QoE). General-purpose processors have evolved to help make
AI performant and cost-effective for the RAN.

In this paper, we’ll introduce several use cases for AI in the radio
access network (RAN):

• Saving energy
• Improving spectrum efficiency
• Improving multiple-input multiple-output (MU-MIMO) efficiency
• Improving resource utilization of O-RAN Distributed Unit
(O-DU) upper PHY
• Improving traffic steering

We’ll highlight demonstrations of these benefits pioneered

by innovative virtualized RAN (vRAN) vendors using
Intel® technologies.

Bringing intelligence to the RAN 4

How AI and machine
learning are used
in RANs
AI/machine learning (ML) improves network efficiency, performance,
resiliency, and reliability and enables new business models.
For example, network behavior can be adjusted dynamically to
respond to changing network demands.

Imagine a situation where user demand occasionally skyrockets

in a specific location, such as at a large concert venue.
AI algorithms could be configured to predict capacity needs
and proactively adjust network resources for increased
bandwidth in the venue during a concert. When the crowds
leave, network resources are reduced in that area or reallocated
to another high-demand area.

How AI and machine learning are used in RANs 5

 These capabilities are powered by new AI prediction, correlation,
and recommendation functions to implement a fully automated
network. Predictive analytics models use data and trends over
time for various use cases, for example, to preemptively optimize
network resources and minimize idle or underutilized resources.

These algorithms can be implemented directly in the network’s

O-RAN Distributed Unit (O-DU), O-RAN Centralized Unit
(O-CU), or in other remote nodes such as the near-real-time
RAN intelligent controller (RIC), non-real-time RIC, or service
management and orchestration, depending on the latency
requirements of the use case.

How AI and machine learning are used in RANs 6

Understanding
the intelligent RAN
architecture
The O-RAN Alliance has defined several use cases
and an architecture for building open, intelligent RANs.
The architecture includes the RIC, as shown in Figure 1.
Alongside the O-RAN Radio Unit (O-RU), O-DU and O-CU,
this architecture includes:

• Near-real-time RIC (near-RT RIC):

This services microservices-based xApps, which can be
used for use cases with a latency tolerance of 10 milliseconds
to 1 second. The near-real-time RIC is typically hosted at
the telco edge or regional cloud.

Understanding the intelligent RAN architecture 7

 • Non-real-time RIC (non-RT RIC):
Hours SMO This services microservices-based rApps, which are used
Service Management and Orchestration
for use cases with a latency tolerance greater than 1 second.
The non-real-time RIC is hosted centrally in the network.
Seconds Non-Real-Time
rApps • Service management and orchestration (SMO):
More than 1 sec RAN Intelligent Controller
The SMO has long been part of the network, but it can now
A1 O1 support the rApps and xApps by processing RAN data,
Milliseconds Near-Real-Time for example, by training AI models.
xApps
10 ms to 1 sec RAN Intelligent Controller
E2 E2 xApps and rApps are typically AI/ML-based models, although
they need not be.
O-RAN Centralized
Milliseconds Unit control plane O-RAN Centralized
10 ms to 1 sec E2 E1 Unit user plane For use cases requiring sub-10 millisecond latencies, AI/ML
F1 - C models must be deployed closer to the edge, such as in the
F1 - U Open
Fronthaul O-DU. In a fully virtualized RAN, AI can augment the functions
Microseconds O-RAN Distributed Unit of Layer 1 or Layer 2 of the RAN stack executing in the O-DU.

Microseconds O-RAN Radio Unit

Figure 1: The O-RAN Intelligent RAN Architecture, showing the different functions, the latency for each one,
and standardized interfaces between them for data collection and policy communication.

Understanding the intelligent RAN architecture 8

The O-RAN Alliance specification defines standard interfaces
to transfer data and communicate policies (also shown in
Figure 1). Real-time data includes such things as signal quality,
telemetry data on the network’s performance, and user
throughput requirements. The scope of data required will
continue to evolve as AI models are developed for both
current use cases and future use cases.

A fully virtualized RAN can most flexibly adapt to evolving data

needs from new use cases. Virtualization allows AI models
to be added to augment or replace existing functions in the
RAN stack pipeline for better performance and efficiency, and
enables new business models. The simple programming model
for general-purpose processors allows rapid and efficient
network evolution, including with AI. AI algorithms can run on the
same hardware as the RAN functions without an accelerator.

In many use cases, applications can use AI to process historical

or current data. The resulting decisions are returned to the RAN
functions nearer the radio as policies, traveling through the
O-RAN standard E2 interface.

Understanding the intelligent RAN architecture 9

Let’s consider MIMO beam management from the non-real-
Location AI RAN application
time RIC. Beam management is a way to target the signal at
a cluster of users. A simple example of this is an office with
Non- • Energy saving (switch cells off/on)
a restaurant below. During the day, the office on the upper
Real- • Radio optimization to improve cell coverage and
floors is busy, while the restaurant on the ground floor is quiet, Time RIC capacity
and at night, the restaurant is full. Beam management would
allocate the beam with the radio frequency (RF) signal quality O-CU • Anomaly detection for RAN key performance
maximized upstairs during the day and downstairs after work. or Near- indicators (KPIs)
Typically, the network could adjust the beam on a schedule Real- • Slice service level agreement (SLA) assurance using
Time RIC an xApp to predict performance indicators and avoid
according to typical user patterns, but the non-real-time
latency violation
RIC would use information from the RAN to optimize beam • Power saving with dynamic CPU frequency (P-state)
management using current demand. • Multi-access traffic steering to improve throughput
• Connection management for handover to improve
Table 1 shows some of the use cases for AI in the RAN and user throughput, load balancing, and coverage
the typical location of the related models.
O-DU • Scheduler parameter optimization for better Quality of
Experience (QoE)
• Link adaptation to improve cell throughput
• User selection to improve spectral efficiency
• Power saving (using C-states to put cores into
microsleep states)
• Channel estimation to improve cell throughput

Table 1: AI use cases in the RAN

Understanding the intelligent RAN architecture 10

Demonstrations of AI
in the RAN
We’ll now explore some demonstrations of AI in the RAN
that Intel has enabled, working with other companies in
the telco ecosystem.

Demonstrations of AI in the RAN 11

 Saving energy through dynamic
power management
The RAN is responsible for 73% of the energy in the mobile
network3, and the O-RU accounts for most of that. Capgemini
and Intel worked together to create an AI-enhanced rApp to
reduce the power used by the O-RU.

The rApp monitors the energy consumption of the RAN

nodes. It then uses a time-series machine-learning algorithm
to forecast future energy consumption, carbon emissions,
and load on the RAN. Input data includes wireless resource
usage, service loads, the number of mobile users, and weather
information. The load prediction uses data from a RAN node
and its neighboring RAN nodes.

The solution was built and optimized using the Chronos

software for scalable time-series analysis, part of the BigDL
framework developed by Intel. The hardware is based on 3rd
Generation Intel® Xeon® Scalable processors.

Demonstrations of AI in the RAN 12

The rApp uses the predicted future load to apply energy-saving
measures. These include:

• Switching off radio sites when they are not needed.

• Switching off high band carriers in a multiband cell when
not required. Users are moved to low-band carriers 3.8 GHz 2.5 GHz 900 MHz 700 MHz
(see Figure 2). 5G/NR 5G/NR 5G/NR 5G/NR

Capgemini tested the solution across twenty 4G and 5G radio

sites with 150 cells. It could switch off up to 120 cells during
the low-traffic period.

The algorithm uses closed-loop automation to improve

the accuracy of the energy-saving decisions over time.

Switched off 2.5 GHz 900 MHz 700 MHz

3.8 GHz 5G/NR 5G/NR 5G/NR 5G/NR

Figure 2: In times of high traffic (top image), all carriers are on. High-band carriers are switched off when traffic
is low, and users are moved to low-band carriers.

Demonstrations of AI in the RAN 13

Improving spectrum efficiency
In another project, Capgemini deployed its NetAnticipate5G
and RATIO O-RAN platform to improve spectrum efficiency
through advanced AI and machine learning (ML) techniques.
This platform forecasts and assigns the appropriate MCS
(modulation and coding scheme) values for signal transmission
by predicting user signal quality and mobility patterns.

In this way, the RAN can intelligently schedule MAC resources

to achieve up to 40% more accurate MCS prediction and
yield up to 15% better spectrum efficiency.4 As a result, it
delivers faster data speeds, better and more consistent QoE to
subscribers, and robust coverage for use cases that rely on low
latency connectivity, such as robotics-based manufacturing
and vehicle-to-everything (V2X) connectivity.

Demonstrations of AI in the RAN 14

 “We gathered and utilized over one terabyte of data and
conducted countless test runs with NetAnticipate5G to
fine-tune the predictive analytics to meet diverse operator
requirements,” said Walid Negm, Chief Research and
Innovation Officer at Capgemini Engineering. “In short,
machine learning can be deployed for intelligent decision-
making on the RAN without any additional hardware
requirement. This makes it cost-efficient in the short run
and future-proof in the long run as we move into cloud-native
RAN implementations.”

Download the white paper: Intelligent 5G L2 MAC Scheduler.

Demonstrations of AI in the RAN 15

 Improving MU-MIMO efficiency
Aira Technologies has demonstrated an xApp to improve
channel estimation and channel prediction performance,
helping to optimize downlink throughput and range.
Accurate channel prediction helps minimize inter-user
interference, which improves multi-user multiple-input
multiple-output (MU-MIMO) performance.

The xApp uses machine learning and is assisted by the Intel®

FlexRANTM reference software, which fully implements O-DU
Layer 1 and above. Intel FlexRAN software enables the xApp
to access the physical Layer 1. The xApp is designed to work in
concert with the VMware RIC platform and uses Capgemini’s
L2 and L3 software platforms. The hardware Aira Technologies
is using is based on Intel® Xeon® processors.

Demonstrations of AI in the RAN 16

“Radio Access Networks (RAN) operate in varied network
conditions and radio environments,” said Ravikiran Gopalan,
Founder and CTO of Aira Technologies. “ML presents a natural
framework to classify these conditions accurately and process
the RAN signals optimally for each of these conditions. We are
seeing tremendous MU-MIMO throughput gains from our
ML-based channel prediction xApp, and we are working on
applying our ML framework to other RAN functionalities.”

Read the Aira press release: xApp improves 5G MU-MIMO

efficiency and throughput using AI.

Demonstrations of AI in the RAN 17

 Improving resource utilization of
O-DU upper PHY software
DeepSig has developed embedded software that replaces
multiple 5G NR signal processing algorithms with a precisely
designed deep neural network (DNN).

This approach potentially requires less computation while

significantly improving network capacity and resilience to
interference by learning the real-world characteristics of
the local wireless environment where the O-RU operates.
These improvements reduce both capital expenditure and
operating expenses. DeepSig and Intel collaborated to bring
this transformational AI software to market as part of the
Intel FlexRAN software suite.

Demonstrations of AI in the RAN 18

DeepSig’s 5G AI embedded software provides drop-in
replacements to the physical uplink shared channel (PUSCH)
channel estimation routines (for standard MIMO), and to SRS
channel estimation and pre-coding routines (for Massive
MIMO). DeepSig’s 5G AI software components can be
leveraged by existing Intel FlexRAN software for O-DU vendors
without additional hardware or software stack changes.

By using machine learning for Layer 1 processing, less compute

is required to process the uplink (PUSCH) signals in standard
MIMO configurations, increasing the number of sectors per
server and reducing their cost of operation.

The ML-driven Layer 1 processing improves the signal-to-

interference-plus-noise ratio (SINR). This enables increased
throughput and coverage, optimizing the value and utilization of
costly spectrum licenses and band allocations. SINR improvements
can also enable bandwidth increases, reduced user traffic
latency, and a smaller interference margin for cell planning.

Download the white paper: Amplifying 5G vRAN performance

with artificial intelligence and deep learning.

Demonstrations of AI in the RAN 19

 Improving traffic steering
Mobility robustness optimization (MRO) dynamically manages
the network performance of handovers to improve end-user
experience and increase network capacity. The aim is to
eliminate radio link failures and reduce unnecessary handovers.
MRO works by automatically adjusting handover boundaries
based on performance indicators. MRO minimizes human
intervention in the network management and optimization
tasks, thus reducing operator expenditures.

Handover decisions made by the MRO capability only consider

cell-level statistics in fine-tuning the handover thresholds.
Traffic management solutions usually relocate users among
cells, relying heavily on user equipment (UE) feedback in the
measurement report. The statistical characteristics of the radio
network and information about the UE behavior are not currently
used to enhance the network and user experience performance.

Demonstrations of AI in the RAN 20

Mavenir has demonstrated the ability to proactively manage
specific user traffic across different cells and access
technologies, using information about the UE behavior and
the radio network. A mobility management rApp provides
policy recommendations for handovers and sends them to
the near-RT RIC for policy enforcement.

The traffic steering xApp hosts a reinforcement learning (RL)

agent. From the O-CU and O-DU, the xApp receives context
and state information about the UE and the serving cells,
the UE’s layer 3 radio resource control (RRC) information
for neighbor cells, and neighbor cell context information.
The RL agent in the xApp then generates a UE-specific
handover control action that optimizes the decision of the
target cell for the UE. This decision maximizes a given key
performance indicator for the UE, such as throughput or
latency. In this way, traffic steering enables operators to
provide an intent-driven UE-specific service.

Demonstrations of AI in the RAN 21

 The outcomes include:

• Better resource utilization of RIC compared to traditional

self-organizing network (SON) and radio resource
management (RRM) algorithms.
• Average 50% improvement in throughput with RIC compared
to SON and RRM.5
• Average 52% improvement in UE spectral efficiency with
RIC compared to SON and RRM.5

The reinforcement learning models to make these RAN

performance gains are accelerated by Intel® Optimization
for TensorFlow, which takes advantage of Intel® AI Engines
for the 4th Generation Intel® Xeon® Scalable processor.
The reinforcement learning AI performance results below
meet Mavenir’s AI performance requirements:

• AI training latency: 7.72ms with single CPU core5

• AI Inference latency: 0.65ms with single CPU core5

Download the white paper: Open vRAN radio intelligent

controller expands RAN Capability.

Demonstrations of AI in the RAN 22

Accelerating AI on
general-purpose
processors
Hardware performance is important for AI-based applications.
The telco network has strict latency requirements for some
functions, and associated AI applications should be processed
close to the wireless transmission. In a virtualized DU, the
same general-purpose processor that supports Layer 1 and 2
functions can also run the AI application.

4th Gen Intel Xeon Scalable processors include several features

that accelerate AI models. These include Intel® Advanced
Vector Extensions (Intel® AVX) for vRAN, support for lower-
precision calculations for faster AI training and inference,
and Intel® Advanced Matrix Extensions (Intel® AMX).

Accelerating AI on general-purpose processors 23

 To help developers take advantage of processor features such
as these, Intel has a strong ecosystem of software support.
Intel has released optimized AI libraries, such as Intel® oneAPI
Deep Neural Network Library, Intel® oneAPI Math Kernel
Library, and Intel® oneAPI Data Analytics Library.

The Intel® oneAPI AI Analytics Toolkit provides tools and

frameworks to accelerate the end-to-end analytics and
AI pipeline on Intel® architecture processors. It provides
an Intel-optimized framework for TensorFlow and PyTorch
with low-precision tools for high-performance training and
inference. It accelerates data preprocessing and machine
learning workflows with the Python packages Modin,
scikit-learn, and XGBoost.

Other AI acceleration features include support for Chronos,

which enables developers to build time-series forecasting
AI models, including data processing and feature engineering.
Intel oneAPI AI Analytics Toolkit includes built-in AI models,
a user-friendly Application Programming Interface (API),
and AutoML for automatic hyperparameter optimization in
a distributed architecture. AutoML enables developers to
train and tune highly accurate AI models.

Demonstrations of AI in the RAN 24

Conclusion
Learn more
While operators have already started deploying AI in select use
cases, there is huge potential, and a number of demonstrations • Amplifying 5G vRAN performance with
have already shown what is possible. artificial intelligence and deep learning

The benefits of introducing AI in several layers in the RAN have • White paper: Intelligent 5G L2 MAC Scheduler
been demonstrated. The flexibility of a fully virtualized RAN • Accelerate AI workloads with Intel® AMX
makes it possible to leverage AI capabilities by enhancing parts
of the RAN processing pipeline.

As operators introduce virtualized RAN to their networks

today, they are laying the foundations for future AI-based
enhancements.

Intel® Xeon® processors include built-in acceleration for

AI training and inferencing and are supported with tools,
frameworks, and software for rapid implementation.
Ecosystem partners, including Capgemini, Aira, DeepSig, and
Mavenir, use them to enable their AI-based RAN technologies.

Conclusion 25
1. European mobile data traffic will triple in next five years, GSMA, November 2023
2. 2023 annual survey highlights, CTIA, November 2023
3. GMSA Intelligence, Going green: Benchmarking the energy efficiency of mobile, June 2021
4. Industry’s first machine learning-based ran application boosts spectral efficiency by 15%, Capgemini, 30 June 2021
5. System: Intel Corporation ArcherCity; Number of nodes: 1 node; Baseboard: Intel Corporation ArcherCity; Chassis: Rack Mount Chassis; CPU Model: Intel® Xeon® Platinum 8480+; Microarchitecture: 4th Gen Intel® Xeon® Scalable Processor; Sockets: 2; Cores per Socket: 56;
Hyperthreading: Enabled; CPUs: 224; Intel Turbo Boost: Enabled; Base Frequency: 2.0GHz; All-core Maximum Frequency: 3.0GHz; Maximum Frequency: 3.8GHz; NUMA Nodes: 2; Prefetchers: L2 HW, L2 Adj., DCU HW, DCU IP; PPINs: 08b7d81fa52198c3,08b0d21f7f5ec0df;
Accelerators: DLB:2, DSA:2, IAX:2, QAT (on CPU):2, QAT (on chipset):0; Installed Memory: 256GB (8x32GB DDR5 4800 MT/s [4800 MT/s]); Hugepagesize: 2048 kB; Transparent Huge Pages: madvise; Automatic NUMA Balancing: Enabled; NIC: 1x Ethernet Controller I225-LM; Disk: 1x
3.6T Samsung SSD 870 QVO 4TB; BIOS: EGSDCRB1.86B.0090.D03.2210040151; Microcode: 0xab000310; OS: Ubuntu 22.04.1 LTS; Kernel: 5.15.0-43-generic; TDP: 350 watts; Power & Perf Policy: Performance; Frequency Governor: performance; Frequency Driver: intel_pstate; Max
C-State: 9; Tensorflow: 2.11; Tested AI Workloads: Mavenir reference learning training and inference for traffic steering xAPP. Tested by Intel as of 04/14/2023

Intel does not control or audit third-party data. You should consult other sources to evaluate accuracy.
Intel technologies may require enabled hardware, software, or service activation.
Your costs and results may vary.
© Intel Corporation. Intel, the Intel logo, and other Intel marks are trademarks of Intel Corporation or its subsidiaries. Other names and brands may be claimed as the property of others. 1223/JH/CAT/PDF 358291-001EN