Source-linked AI summary
AI-RAN: Transforming RAN with AI-driven Computing Infrastructure
Lopamudra Kundu, Xingqin Lin, Rajesh Gadiyar, Jean-Francois Lacasse, Shuvo Chowdhury
TL;DR
AI-RAN addresses the shift from communication-centric RANs toward converged infrastructures that host both RAN and AI workloads. The article organizes AI-RAN’s forms, identifies convergence requirements, proposes a reference architecture, and demonstrates concurrent processing on shared hardware. The proof-of-concept shows improved GPU utilization when RAN and AI workloads run together, while the architecture remains an early-stage area for future development.
Problem
Existing work lacks a unified framework for the diverse uses of AI in RAN as communication and computing converge.
Method
The article categorizes AI-RAN forms, defines requirements and enablers, presents a reference architecture, and demonstrates concurrent RAN and AI processing.
Results
RAN and AI workloads run concurrently on shared infrastructure, with GPU utilization rising from a RAN-only peak of 40% by using remaining resources for AI workloads.
Takeaways & Limitations
AI-RAN supports adaptive resource allocation and shared infrastructure for more efficient and intelligent networks serving users, services, and vertical industries.
Abstract
from arXiv · showhide
The radio access network (RAN) landscape is undergoing a transformative shift from traditional, communication-centric infrastructures towards converged compute-communication platforms. This article introduces AI-RAN which integrates both RAN and artificial intelligence (AI) workloads on the same infrastructure. By doing so, AI-RAN not only meets the performance demands of future networks but also improves asset utilization. We begin by examining how RANs have evolved beyond mobile broadband towards AI-RAN and articulating manifestations of AI-RAN into three forms: AI-for-RAN, AI-on-RAN, and AI-and-RAN. Next, we identify the key requirements and enablers for the convergence of communication and computing in AI-RAN. We then provide a reference architecture for advancing AI-RAN from concept to practice. To illustrate the practical potential of AI-RAN, we present a proof-of-concept that concurrently processes RAN and AI workloads utilizing NVIDIA Grace-Hopper GH200 servers. Finally, we conclude the article by outlining future work directions to guide further developments of AI-RAN.
I. INTRODUCTION
RANs are evolving from dedicated mobile-broadband infrastructure toward converged platforms that host communication and computing workloads. The article addresses the lack of a unified framework for AI’s diverse RAN uses by organizing AI-RAN manifestations and presenting requirements, architecture, demonstrations, and future directions.
- I. INTRODUCTION: RANs are expanding beyond mobile broadband to support edge computing, vertical use cases, and heterogeneous workloads.This evolution brings communication and computing services together at the network edge.
- I. INTRODUCTION: AI-RAN runs RAN and AI workloads on shared edge infrastructure, including support for AI-as-a-service.Shared infrastructure can dynamically allocate resources across workloads and improve asset utilization.
- I. INTRODUCTION: AI-RAN research also targets AI-native functionality embedded in PHY and MAC protocols for future air interfaces.The convergence of communication and computing is presented as essential for these capabilities.
- I. INTRODUCTION: Existing work lacks a unified framework covering AI’s diverse uses in RAN.The article responds with use-case forms, requirements and enablers, a reference architecture, proof-of-concepts, and future work.
II. USE CASES OF AI-RAN
AI-RAN manifests across protocol, service, and infrastructure levels as AI-for-RAN, AI-on-RAN, and AI-and-RAN. These forms improve RAN functions, host AI applications, and share computing resources between AI and RAN workloads.
- AI-for-RAN: AI-for-RAN applies AI across the RAN protocol stack to improve spectral efficiency, operating cost, user experience, and network automation.Examples span PHY functions such as channel estimation and beamforming, MAC functions such as scheduling, and zero-touch management.
- AI-on-RAN: AI-on-RAN uses RAN infrastructure to host AI applications and AI-as-a-service for internal teams, enterprises, and industry verticals.Examples include smart-city and IoT applications, GenAI tools, LLMs, and personalized customer care services.
- AI-and-RAN: AI-and-RAN dynamically shares computing resources between RAN network functions and AI applications on common infrastructure.The approach supports concurrent workloads, multi-tenancy, and resource repurposing according to demand.
- AI-and-RAN: 30%-40% to nearly 100%: AMCOP reports increasing hardware utilization by running AI and RAN workloads in parallel instead of RAN alone.The passage attributes the change to monitoring unused GPU cycles and dynamically allocating them to AI applications.
III. REQUIREMENTS AND ENABLERS OF AI-RAN
AI-RAN requires coordinated computing and communication capabilities, including accelerated hardware, cloud-native software, joint orchestration, and native AI support. These enablers provide the foundation for customizable AI-RAN deployments.
- Requirements and enablers: AI-RAN must combine high-performance computing and communication functions while supporting scalability and intelligent operation.The design identifies requirements and technology enablers for realizing this convergence.
- Accelerated computing infrastructure: General-purpose accelerators such as GPUs are needed for intensive RAN and AI workloads that traditional ASIC and CPU infrastructure cannot handle.Massively parallel processing and high-speed, low-latency interconnects support workloads including MIMO beamforming and scheduling.
- Software-defined, cloud-native design: Software-defined, cloud-native design abstracts network functions from hardware and supports flexible, portable, programmable, and elastic deployments.Containers and microservices enable resource-efficient multi-tenancy and scaling according to demand.
- Joint orchestration: Joint orchestration allocates computational and communication resources on demand according to traffic, user demand, workload requirements, and SLAs.The orchestrator can place latency-sensitive tasks at the edge and route compute-heavy non-real-time tasks centrally, while scaling resources as conditions change.
- Native AI support: A network digital twin provides a virtual replica for data collection, simulation, training, and validation before AI models are deployed in the live network.This sandbox helps test model behavior without risking unintended issues such as downtime.
- Requirements and enablers: Combining these features establishes a foundation on which operators can build customized AI-RAN deployments.The foundation coalesces the technology enablers into an architecture for deployment.
IV. REFERENCE ARCHITECTURE DESIGN
The reference architecture provides a software-defined, accelerated datacenter blueprint for colocating AI and RAN workloads. It combines AI-RAN servers, scalable Ethernet fabrics, and a cloud-native software stack to support multi-tenancy and deployment across use cases.
- IV. REFERENCE ARCHITECTURE DESIGN: The architecture uses datacenter racks containing AI-RAN servers with CPUs, GPUs, DPUs, and SSDs connected through Ethernet switching fabric.It is intended as a scalable blueprint for next-generation software-defined and accelerated AI-RAN datacenters.
- Network topology: The compute fabric connects radio units to AI-RAN servers through aggregation routers and a scalable two-spine, four-leaf topology.Fronthaul leaf switches serve radio-unit access, while server leaf switches connect the spine layer to AI-RAN servers.
- Network topology: A converged fabric connects AI-RAN servers toward midhaul, backhaul, or the internet according to which RAN functions are hosted.The connection varies for deployments hosting DU, DU and CU, or DU, CU, and CN combinations.
- Software stack: The software stack uses a cloud operating system such as Kubernetes and APIs such as CUDA to orchestrate resources and manage infrastructure.The stack supports AI and RAN multi-tenancy on the same platform while maintaining desired RAN service requirements.
- Deployment blueprint: The resulting architecture and software stack form a deployment blueprint for enabling AI with RAN on shared infrastructure.The article uses a proof-of-concept to demonstrate one supported use case.
V. AI-RAN PROOF-OF-CONCEPT
The proof of concept demonstrates concurrent RAN and AI execution on shared GPU infrastructure, using workload partitioning to improve utilization and reduce idle capacity.
- Concurrent RAN and AI workloads run on the same server through time- or space-based multi-tenancy.An orchestrator provisions, deprovisions, or shifts workloads according to available capacity and changing compute needs.
- The experiment serves a chatbot or digital human with an LLM while the same server processes 5G RAN traffic.
- Concurrent AI and 5G operation improves GPU utilization and minimizes idle states compared with a RAN-only workload peaking at 40% usage.The resulting capacity may reduce provider total cost of ownership by supporting additional services when demand permits.
VI. CONCLUSION AND FUTURE WORK
AI-RAN converges communication and computing by running RAN and AI workloads on shared infrastructure, supporting more adaptive resource allocation. The article concludes that the field remains early-stage and identifies orchestration, interoperability, and real-world validation as priorities.
- AI-RAN shifts RANs toward converged compute-communication platforms that support AI-native functionalities.
- Shared infrastructure allows AI-RAN to dynamically allocate resources across users, services, and emerging vertical industries.
- AI-RAN architectures and frameworks remain in their early developmental stages.
- Future work: Future orchestration frameworks should manage compute and communication resources while accounting for latency, reliability, and energy constraints.
- Future work: Open interfaces and data formats are needed to support interoperable multivendor AI-RAN solutions.
- Future work: Testbeds, benchmarks, simulations, experiments, and pilot deployments are needed to validate solutions and refine best practices.