Description

1. Solution Overview

The AI-RAN Aerial Testbed Solution is a complete, GPU-accelerated 5G and AI-RAN development platform designed for universities, research institutes, mobile network operators, equipment vendors, and wireless engineering teams.

Based on the NVIDIA Aerial Testbed 1.0 architecture, the solution integrates NVIDIA Aerial CUDA-Accelerated RAN, OpenAirInterface, O-RAN-compliant radio units, precision timing equipment, fronthaul networking, a 5G Core Network, and commercial user equipment into a fully operational end-to-end wireless test environment.

The platform implements the O-DU Low physical layer on an NVIDIA GPU, while OpenAirInterface provides the O-DU High, O-CU, and 5G Core Network functions. It supports both over-the-air and RF-cabled operation and can interoperate with commercial O-RUs and 5G user devices through an O-RAN 7.2x fronthaul interface. (NVIDIA Docs)

The system is suitable for:

• 5G and 6G physical-layer research
• AI-RAN algorithm development
• O-RAN interoperability testing
• Neural receiver and AI-enhanced PHY research
• RAN scheduling and resource-allocation research
• Real-time wireless data collection
• Digital-twin-assisted network validation
• Commercial UE and O-RU integration
• End-to-end application and throughput testing

---

2. Reference System Architecture

The Aerial solution provides a complete end-to-end network chain:

Commercial UE → O-RU → O-RAN 7.2x Fronthaul → Fronthaul Switch → NVIDIA Aerial O-DU → OAI O-DU High/O-CU → 5G Core Network → Application Server

The reference architecture combines:

• NVIDIA GPU-accelerated Layer 1
• OpenAirInterface Layer 2 and Layer 3
• O-RAN Category A radio units
• Precision Time Protocol synchronization
• Commercial 5G smartphones or UE emulators
• Real-time RAN data acquisition
• AI/ML training and inference workflows

The O-RU connection uses Ethernet-based eCPRI transport over an O-RAN 7.2x interface, including Control Plane, User Plane, Synchronization Plane, and Management Plane support.

---

3. Hardware Configuration

3.1 AI-RAN DU/CU Computing Platform

Recommended Grace Hopper Configuration

Supermicro ARS-11GL-NHR Grace Hopper Server

Component	Specification
CPU	72-core NVIDIA Grace CPU
GPU	NVIDIA GH200 Grace Hopper Superchip
GPU Architecture	NVIDIA Hopper
System Memory	480 GB LPDDR5X with ECC
GPU Memory	Up to 96 GB HBM3
Network Interfaces	Two NVIDIA BlueField-3 DPUs/NICs
Storage	Enterprise NVMe SSD
Operating Architecture	ARM64
Primary Role	O-DU Low, O-DU High, O-CU and optional 5G Core

The Grace Hopper architecture provides a high-bandwidth coherent connection between the CPU and GPU, allowing Layer 1 workloads, Layer 2 processing, AI inference, and data acquisition to operate on a tightly integrated computing platform.

The official Aerial Testbed bill of materials specifies the Supermicro ARS-11GL-NHR with a 72-core Grace CPU, GH200 GPU, 480 GB ECC LPDDR5X memory, and two BlueField-3 network interfaces.

Compact Configuration

NVIDIA DGX Spark

Component	Specification
CPU	20-core ARM CPU
CPU Architecture	10 Cortex-X925 cores and 10 Cortex-A725 cores
GPU	NVIDIA GB10 Blackwell architecture
Unified Memory	128 GB
Network	Two QSFP interfaces based on ConnectX-7
Primary Role	Compact single-cell Aerial Testbed

Aerial Testbed 1.0 adds support for implementing an over-the-air testbed on DGX Spark. The reference release supports single-cell operation on this platform.

---

3.2 Fronthaul Network Switch

Recommended Model

Dell PowerSwitch S5248F-ON

The fronthaul switch provides deterministic, high-bandwidth connectivity among the NVIDIA computing platform, O-RU, and PTP Grandmaster.

Typical functions include:

• 25GbE O-RU fronthaul connectivity
• 100GbE or 200GbE server connectivity
• IEEE 1588v2 Precision Time Protocol
• VLAN-based fronthaul isolation
• Jumbo-frame transmission
• Low-latency Ethernet switching
• O-RAN synchronization-plane transport

Other switch platforms listed in the NVIDIA reference bill of materials include:

• NVIDIA Spectrum SN3750
• NVIDIA Spectrum SN5400
• Cisco Nexus 9300
• Ciena 5164
• ADVA XG480
• FibroLAN Falcon RX

---

3.3 PTP Grandmaster

VIAVI Qg 2 Multi-Sync Gateway

The VIAVI Qg 2 provides precision frequency and phase synchronization for the O-DU, fronthaul switch, and O-RU.

Primary functions include:

• IEEE 1588v2 PTP Grandmaster
• GNSS-referenced timing
• Frequency synchronization
• Phase and time synchronization
• Telecom timing profile support
• Synchronization monitoring
• Multi-device timing distribution

Accurate synchronization is essential for O-RAN 7.2x TDD operation, coordinated radio transmission, and deterministic fronthaul processing. The VIAVI Qg 2 is the Grandmaster identified in the official Aerial Testbed hardware list.

---

3.4 O-RAN Radio Unit

Reference O-RU

WNC 1220 4T4R Indoor O-RU

Parameter	Specification
Radio Architecture	4T4R
O-RAN Category	Category A
Functional Split	O-RAN 7.2x
Fronthaul Interface	25GbE
Transport	Ethernet/eCPRI
Reference Frequency Range	3.55–3.77 GHz
Deployment Type	Indoor
Synchronization	PTP over O-RAN S-Plane

The platform may also be integrated with alternative n77, n78, n79 or other sub-6 GHz O-RUs, subject to band selection, firmware compatibility and O-RAN interoperability validation.

The official Aerial Testbed documentation identifies the WNC 1220 as the qualified 4T4R reference O-RU and lists n48 and n78 as tested sub-6 GHz reference bands for the platform.

---

3.5 5G User Equipment

Recommended commercial user equipment includes:

• Samsung Galaxy S22
• Samsung Galaxy S23
• Engineering 5G CPE
• Commercial 5G smartphones
• OAI GPU-accelerated Soft-UE
• External UE simulator

The Samsung Galaxy S22 and S23 reference devices support up to four downlink SU-MIMO layers and one uplink layer in the qualified Aerial Testbed configuration.

---

3.6 5G Core Network

The 5G Core Network may run:

• On the same Grace Hopper server as the gNB
• On a separate ARM-based server
• On a separate x86-based server
• In containerized or cloud-native form

Core functions include:

• Access and Mobility Management Function
• Session Management Function
• User Plane Function
• Authentication Server Function
• Unified Data Management
• Unified Data Repository
• Network Repository Function

The same Supermicro Grace Hopper server can host both the gNB and the Aerial Testbed 5G Core, although a separate Core Network server may be used when required.

---

3.7 Accessories

The complete solution normally includes:

• QSFP28 direct-attach cables
• 25GbE optical transceivers
• 10GbE optical transceivers
• LC-to-LC multimode optical cables
• Ethernet management cables
• RF attenuators and RF cables
• O-RU power supply
• Rack-mounted PDU
• GNSS antenna and timing cables
• SIM cards and programmable SIM profiles
• Management workstation
• Keyboard, mouse and display accessories
• Rack installation components

---

4. Software Stack

NVIDIA Aerial CUDA-Accelerated RAN

The GPU-accelerated Layer 1 stack provides:

• Downlink and uplink 5G NR physical-layer processing
• GPU-accelerated channel coding and decoding
• LDPC processing
• Polar-code processing
• FFT and IFFT processing
• Channel estimation
• Equalization
• Modulation and demodulation
• MIMO processing
• Beamforming
• PRACH processing
• PDSCH and PUSCH processing
• PDCCH and PUCCH processing
• O-RAN fronthaul processing
• FAPI-based Layer 1/Layer 2 integration

Aerial provides full inline acceleration of physical-layer functions and selected MAC functions on NVIDIA GPUs.

OpenAirInterface

OpenAirInterface provides:

• O-DU High
• MAC and RLC
• O-CU-CP
• O-CU-UP
• RRC and PDCP
• 5G Core Network
• Commercial UE registration
• PDU session establishment
• End-to-end IP connectivity

Containerized Deployment

The software stack can be delivered as validated containers for:

• NVIDIA Aerial Layer 1
• OAI gNB Layer 2 and Layer 3
• OAI 5G Core
• Network management
• Monitoring and logging
• Data Lake services
• AI/ML applications

---

5. Main Functions

5.1 End-to-End 5G Standalone Network

The system provides a complete 5G SA network from the radio interface to the Core Network. Commercial UEs can register with the network, establish PDU sessions and run bidirectional IP traffic.

Supported test modes include:

• Over-the-air testing
• RF-cabled testing
• Shielded-box testing
• UE-emulator testing
• O-RU-emulator testing
• Soft-UE testing

---

5.2 GPU-Accelerated Physical Layer

The complete 5G NR Layer 1 processing chain is executed inline on the NVIDIA GPU.

Benefits include:

• High-throughput signal processing
• Deterministic real-time execution
• Lower CPU processing requirements
• Scalable cell processing
• Direct integration of AI and RAN workloads
• Rapid implementation of custom PHY algorithms

---

5.3 O-RAN 7.2x Integration

The platform implements an O-RAN 7.2x fronthaul architecture between the O-DU and O-RU.

It supports:

• C-Plane transmission
• U-Plane transmission and reception
• S-Plane synchronization
• M-Plane radio management
• eCPRI encapsulation
• 25GbE O-RU connectivity
• Single-cell and multi-cell research
• Multi-vendor O-RU interoperability

The reference Category A split places precoding, digital beamforming, IFFT and cyclic-prefix insertion in the downlink low-PHY, and FFT, cyclic-prefix removal and digital beamforming in the uplink low-PHY.

---

5.4 AI-RAN Algorithm Development

Researchers can integrate AI models into different sections of the RAN protocol stack.

Typical research areas include:

• Neural channel estimation
• Neural receivers
• AI-based equalization
• Learned demodulation
• Link adaptation
• Intelligent scheduling
• Traffic prediction
• Energy optimization
• Interference management
• Beam selection
• Spectrum sensing
• Integrated sensing and communications
• Deep reinforcement learning for MAC scheduling

Because Aerial and OAI expose modifiable physical-layer and Layer 2 software, researchers can customize modulation, coding, signal-processing and scheduling algorithms.

---

5.5 Real-Time dApp Framework

Aerial Testbed 1.0 expands support for distributed applications, or dApps.

Applications may be developed using:

• Python
• C
• CUDA
• PyTorch

A dApp can subscribe to real-time RAN information such as:

• Fronthaul time-frequency resource-block I/Q samples
• Channel-state information
• PUSCH pipeline signals
• Channel estimates
• Layer 1 and Layer 2 metadata

This allows real-time AI inference for spectrum management, sensing, radio optimization and other AI-RAN functions.

---

5.6 Aerial Data Lake

The Aerial Data Lake enables synchronized capture of RAN data for offline analysis and real-time AI applications.

Supported data includes:

• Uplink fronthaul I/Q samples
• PUSCH data
• Channel estimates
• FAPI messages
• Layer 1 metadata
• Layer 2 scheduling information
• Downlink transport blocks

The captured data can be used for:

• AI/ML training dataset generation
• Neural-receiver development
• Troubleshooting
• RF-performance analysis
• Algorithm benchmarking
• Real-time dApp inference

The Data Lake provides time-coherent capture and database APIs for retrieving collected RAN data.

---

5.7 Digital Twin and Channel Emulator Integration

The platform can be connected to:

• NVIDIA Aerial Omniverse Digital Twin
• RF channel emulators
• Commercial UE emulators
• O-RU emulators
• Site-specific ray-tracing models
• Standardized 3GPP stochastic channel models

This enables repeatable validation under realistic propagation conditions before field deployment.

---

6. Performance

Reference OTA Performance

On a Grace Hopper-based Supermicro platform, previously published Aerial Testbed validation results include:

Performance Item	Reference Result
Peak Downlink Throughput	Approximately 1.03 Gbps
Peak Uplink Throughput	Approximately 125 Mbps
Two-Cell Downlink Throughput	Approximately 700 Mbps per cell
Two-Cell Uplink Throughput	Approximately 105 Mbps per cell
Maximum Downlink MIMO Layers	Four layers
RF Channel Bandwidth	Up to 100 MHz, depending on configuration
Fronthaul Interface	25GbE per reference O-RU
Functional Split	O-RAN 7.2x

These throughput figures were published for the Supermicro Grace Hopper Aerial platform in NVIDIA’s earlier ARC-OTA validation results. Actual throughput depends on the O-RU, UE capability, frequency band, TDD pattern, MCS, RF conditions, software release and system configuration.

Aerial Platform Scalability

The qualified NVIDIA Aerial CUDA-Accelerated RAN platform supports:

• Up to 20 peak-load 4T4R cells on a Grace Hopper MGX system
• Up to 20 average-load 4T4R BFP9 cells
• 64T64R Massive MIMO
• 100 MHz channel bandwidth
• Up to 16 downlink and 8 uplink Massive-MIMO processing layers
• SRS-based beamforming

These figures represent Aerial CUDA-Accelerated RAN platform capacity and should not be interpreted as the default out-of-the-box capacity of every Aerial Testbed 1.0 OTA configuration.

---

7. Delivery Scope

A complete turnkey Aerial solution can include:

• All required AI-RAN computing hardware
• NVIDIA BlueField-3 networking
• O-RAN fronthaul switch
• PTP Grandmaster
• 4T4R O-RU
• Commercial 5G UE
• SIM cards
• Optical modules and cables
• NVIDIA Aerial software installation
• OpenAirInterface installation
• 5G Core Network deployment
• O-RU configuration
• Switch and PTP configuration
• End-to-end system integration
• OTA or RF-cabled validation
• Throughput testing
• System documentation
• Administrator and developer training
• Remote technical support
• Optional on-site installation

---

8. Key Advantages

Turnkey End-to-End Platform
The complete gNB, O-RAN fronthaul, O-RU, 5G Core and UE environment is integrated and validated before delivery.

GPU-Native RAN Architecture
Real-time physical-layer processing and AI workloads can operate on a common NVIDIA accelerated-computing platform.

Open and Programmable
The Aerial and OpenAirInterface software stacks allow researchers to inspect and modify PHY, MAC, scheduler and Core Network functions.

Commercial Hardware Interoperability
The system supports real O-RUs and commercial 5G user equipment instead of relying exclusively on simulation.

AI-Ready Data Pipeline
Aerial Data Lake and the dApp framework provide direct access to real-time wireless data for model training and inference.

Scalable from Research to Deployment
The architecture supports compact DGX Spark testbeds, full Grace Hopper research platforms, multi-cell validation and high-capacity AI-RAN development.

---

Product Summary

The AI-RAN Aerial Testbed Solution provides a production-relevant environment for developing, validating and demonstrating advanced 5G, O-RAN and AI-native wireless technologies. By combining NVIDIA GPU acceleration, an open software-defined RAN stack, commercial O-RAN radio units, precision synchronization and real-time wireless data acquisition, the platform shortens the path from simulation and algorithm development to live-network validation.