S52156 A bridge to 6G - Aerial Research And Innovation Platform

Florian Kaltenberger, Professor

Chris Dick, Architect Anupa Kelkar, Product Manager
[email protected]
[email protected] [email protected]
(EURECOM / OSA)
(NVIDIA) (NVIDIA)
S52156 A bridge to 6G - Aerial Research And
Innovation Platform

NVIDIA and OAI experts provide an introduction to

the first fully programmable Advanced 5G+ network
as a sandbox – full-stack democratized platform for
all researchers to simulate-prototype-benchmark
optimizations, algorithms, and innovations rapidly in
a deployed over-the-air NR standards compliant high
performant operational network. This session will
highlight platform vision, early adopter use cases,
showcase complete C/C++ programmability, provide
OAI ISV gNB and CN overview and deep dive into
specific ML examples to jumpstart NG innovations.
 Agenda
• Aerial Research Cloud Product and Early Adopters

• OpenAirInterface Software Alliance - ISV Overview

• Future with Aerial Research Cloud

• Summary

•
Aerial Research Cloud
Product and Early Adopters

Anupa Kelkar, Product Manager

(NVIDIA)
 Summary
Background Context

• Time to develop 6G is Now

Globally 5G deployment is in full swing by traditional telco
and private enterprises – it is time to innovate towards 6G

• Adoption of Open RAN architecture

Complete network flexibility and inter-operability with
SIM 5G hardware and software disaggregation and virtualization
SUBSCRIBERS CONNECTIONS CONNECTIONS

2021, 5.3 billion 67% 2025 projected SIM 2025 Advanced 5G • GPU accelerated compute and ML
global population connections 23.3
billion
NR >25%
connections GPU high performance compute, high bandwidth IO, ML
toolchains are necessary to meet 6G requirements

GLOBAL
ECONOMY
NVIDIA – Building the wireless platform for NG
2035 5G+/6G enables >
$10 trillion of global
economic output
[1] Reference: GSMA Mobile Economy 2022
[2] Reference https://www.123rf.com
 What Problems We are Solving
Key Challenges

Lack of 5G OTA
standards compliant
flexible & scalable Functionality cannot
Weeks, days, months Several custom and wireless platforms be changed based on
to prototype expensive use cases or
black boxes + research goals
ML tool chains,
datasets, models

IN-EFFICIENT MULTI BOX PLATFORM NOT NON

SLOW PROCESSESINFRASTRUCTURE READILY AVAILABLE PROGRAMMABLE
DESIGN FLOW

Innovation in advanced ML based wireless research and new radio development hindered by lack
of 3GPP compliant OTA programmable network as a platform
 Next Generation Wireless Research Market Survey
Advanced 5G and 6G Top Researcher Requirements

1. Easy access to high performance cloud native accelerated compute resources

2. Standards compliant OTA network for optimal real time closed loop experimentation

3. Ease of programmability and customization of the network (CN + RAN + UE)

4. Distributed AI, Online training, Federated learning, Network Prediction, Native ML Sim

5. Access to large scale datasets, digital twins and other simulation environments

6. Radios that support 6G spectrum (mmWave) and advanced technologies (massive MIMO)

NVIDIA GPU high performance compute and high BW IO meet 6G requirements

 Meet Aerial Research Cloud
Launchpad to accelerate innovations in advanced wireless communications

PROGRAMMABLE MACHINE LEARNING

5G FULL-STACK

CUDA

GPU + NIC

Unique ecosystem opportunity to usher innovations in advanced wireless rapidly with

machine learning + programmable 5G full-stack platform
 Aerial Research Cloud – Value Pillars
First fully programmable and extensible Advanced 5G network

• Access to source (C/C++)

1
• Easy to extend and change network
ARC Programmable behavior from L1 – CN
Aerial Research Cloud
• Key Network Element with GPU HPC
2 • ML toolchains, MIG architecture
ML ready • Complex compute offload, cell-less

1 2 3 4 3 • Disaggregated hardware and cloud

native software from COTS vendors
Open RAN
• GPU co-host, inline accelerated
Blueprint
First Programmable Network 4
• Advanced 5G NR network component
Network As A BOM - NVIDIA lab qualified
Platform

NVIDIA Aerial Research Cloud (ARC) - built grounds up to enable 6G

 NVIDIA AERIAL RESEARCH CLOUD
SOLUTION ARCHITECTURE
NVIDIA AERIAL RESEARCH CLOUD SOLUTION ARCHITECTURE
Flexible, Scalable, Extensible Deployment Topology View

O-RAN FH 7.2 Cloud Native 5G NR RAN Cloud Native 5G Core On Prem Data Center or Cloud Applications
GPS
Fronthaul Switch
RJ45

100G SFP 10G SFP

GrandMaster
10G SFP gNB CN

vCU
Radio Unit
PCF
RRC
UDM AUSF
PDCP

3GPP
F1 F1 N1 / N2 SMF
vDU AMF

RLC
N4
UPF
MAC
N3 N6
SCF FAPI

ASDK L1
SIM User
Profiles ORAN 7.2 FH

Easy to setup, integrate, deploy and operate

10
 Aerial gives us the flexibility to
develop next-generation wireless
“
systems and applications
based on AI/ML
“
— Tommaso Melodia, Institute for the Wireless Internet of Things at Northeastern University
Research Focus: Dynamic optimization of
advanced cellular networks for industrial
control
AI-based channel
and interference
estimation

slicing and dynamic

resource allocation

X-Mili
A programmable O- Load balancing and
RAN testbed for 5G mobility management
and beyond at
Northeastern
Northeastern University is
deploying a next-generation cellular
testbed with 8 nodes using the Northeastern lab
NVIDIA Aerial Research Cloud and factory floor
the OpenRAN Gym O-RAN
framework
 Aerial provides an exciting
opportunity to demonstrate the
“
power of AI in enhancing the
performance and capabilities of 5G
networks and beyond. “
— Ashutosh Sabharwal, Professor & Department Chair at Rice ECE.
Research Focus: AI enhanced MIMO operation Null
Local

App
Scheduler
App

Bandwidth
Time
Aerial App
Layers Radar Detection and
Compute Footprint Co-existence

Research at Rice
University Centralized Local
App
Scheduler Scheduler

App
Aerial enables several research
directions – near term use cases CPU/GPU Aerial
include enhancement of MIMO
operation using AI. Compute Footprint
Beam Selection and Tracking
Rice Wireless group has been pioneering
the massive MIMO research and
experimental platform design in the last
Local
two decades. Scheduler
App

Compute-Aware Reinforcement
learning-based Resource Scheduling Aerial

Compute Footprint

Deep Learning-based Massive MIMO

Edge Cloud Data Center Detection
 Aerial Research Cloud Roadmap
Timeline View

2023 – Jan. 2024 – Jan. 2025 – Jan.

Aerial Research Cloud CBRS FR2
100Mhz 4T4R n78 3GPP Rel 15 Sub-6 United Stated Coverage Integrated FR2 support for NG
and ORAN 7.2 compliant lab research
network

2023 – Oct. 2024 – Oct.

RIC and ML 32T32R MU-MIMO
OSC RIC Infrastructure and Multi-cell 100Mhz 32T32R MU-
sample Apps MIMO support

C/C++ programmable full-stack to ease developer community extensions

 WHAT IS IN IT FOR YOU?

TRADITIONAL Multi-Vendor Access to Source ML Ready

Proprietary HW/SW (ASIC, FPGA) No fixed function HW Change, Extend, Innovate Off the shelf (GPU, NIC, CPU)
(Limited knobs) (Open Interfaces) (Network Sandbox) (Early adopter extensions)

Innovate at the link layer

Code algorithm in Sionna

GPU accelerated simulation
• Multi-GPU support
• Fast BLER charts

Integrate with Aerial L1

Test, benchmark on ARC

Implement on real network

Proprietary Off the shelf Sionna First fully ML Ready

Network Equipment O-RAN aligned compute programmable (Toolchains like
network Pytorch)

Complete network is Rapidly Simulate- Democratize access to

Software-Defined cloud Network As A Sandbox
programmable in C/C++ Prototype-Benchmark datasets, ML models
native real-time network Full-Stack BOM
innovations

Simulate
Benchmark

2023 Early Adopter

Democratized 5G+ ML
5G NR Launchpad
RAN spending for
$18
[1] http//www.o-ran.org/
[2] Images are Copyright by O-RAN Alliance
Deployments Prototype
Advanced Wireless
billion in 2021
Aerial Research and Innovation Platform Integration
With OpenAirInterface 5G Software

Florian Kaltenberger,
Professor (Eurecom/OSA)
5G Networks are becoming Virtualized, Open and Flexible
O-RAN Blueprint

Radio Network Edge Cloud Core Cloud

O-RU O-CU-C AMF SMF

F1-C NG-C

E1
E2
O-DU O-CU-U UPF
O-RAN 7.2 F1-U NG-U

O-RU
Access and Mobility Management Function (AMF)
E2 Near Real- E2 Session Management Function (SMF)
time RIC User Plane Function (UPF)
O-RAN Centralized Unit (O-CU),
O-RAN Distributed Unit (O-DU),
RAN Intelligent Controller (RIC),
O-RAN Radio Unit (O-RU),
 OpenAirInterface
Democratizing innovation in the 5G era

• The most complete, open-source implementation of 3GPP 4G/5G Radio Access and Core
Network
• Compatible with common-off-the-shelf (COTS) equipment
• Supports non-standalone and standalone 5G deployments
• Complete protocol stack ready for cloud-native deployment on x86 architecture
• Follows open RAN architecture and interfaces specified by O-RAN Alliance
• Founded by EURECOM and managed today by the OpenAirInterface Software Alliance (OSA)
• Community driven development supported by OSA
• The OAI public license is based on Apache 2.0 but allows patent owners to contribute and
keep their patents under 3GPP FRAND rules

www.openairinterface.org
 Integration of Aerial with OpenAirInterface
Layer 1 of DU is by far the most computationally expensive part of a virtualized open RAN deployment!

3GPP NG
• Nvidia Aerial is used as an inline L1 accelerator
• DU-high (Layer 2) from OpenAirInterface
CU SDAP/ RRC/
PDCP PDCP • Interface realized using 5G FAPI defined by Small Cell
Forum
3GPP F1
• Both run in individual docker containers
DU-high • Traces can be analyzed with Wireshark
RLC

MAC • DU Layer 1 southbound interface: O-RAN 7.2

• DU Layer 2 northbound interface: 3GPP F1
SCF 5G FAPI
• DU can be combined with CU in a single
DU-low executable/container
PHY

O-RAN 7.2
Inline GPU acceleration of the entire L1 improves
performance and extends features of OAI DU to
O-RU multiple cell deployments and other advanced
configurations
 Aerial Research Cloud Setup
Example Configuration

COTS UE
5GC
COTS Server O-RU
SMF

UPF AMF
PTP
grandmaster
NG with GPS
SDAP/ RRC/
COTS Server
CU
PDCP PDCP

DU-high MAC/RLC
A100 GPU + CX6-DX NIC
SCF 5G FAPI

DU-low PHY O-RAN 7.2

Fronthaul
switch

Demo available at O-RAN virtual exhibition (www.virtualexhibition.o-ran.org)

 Aerial Research Cloud Workflow
Simple workflow to enable rapid prototyping

1. Start the containers (OAI core network, OAI

CU/DU, Nvidia Layer 1)
2. Start the UE and analyze the connection
procedure in real-time
3. Launch an application on UE and/or on the edge-
cloud
4. Disconnect UE, Stop containers
5. Analyze logs offline (e.g. FAPI messages)
 Key Takeaways

• OpenAirInterface provides a complete open-source implementation of 3GPP 5G

• Aerial Layer 1 acceleration makes scaling up easier and adds features
• Aerial Research Cloud is ideal basis for experimental 6G research
• OpenAirInterface is publicly available from Eurecom’s gitlab server
• You can contribute to the project as a developer or support the project by joining the OpenAirInterface Software
Alliance: https://openairinterface.org/
Working with the Aerial Research Cloud
 ARC – Enabling the Future of Wireless
Research community extending ARC stack and doing AI/ML research with ARC

• AI/ML is critical for B5G and the foundation of 6G

• The industry is in the innovation and benchmarking phase of this work

• Simulation remains critical but need to characterize AI/ML approaches on real OTA systems and networks

• NVIDIA ARC is providing new tools to enable a workflow from algorithm prototyping to OTA
• Native AI/ML link layer simulator
• OTA network testbed

• The research community

• Is extending the ARC software stack to bring new capabilities
• Is using ARC for AI/ML research
• Is exploiting the GPU for full stack acceleration beyond L1

• Let’s look at some examples

• NEU integration of OSC RIC into ARC stack
• Joint work with ETH on ML in L1
• ARC L2 source and acceleration example
 From Algorithm Innovation to OTA

Innovating at Concept
From Sionna
AI/ML algorithm design
Differentiable GPU Ray Tracing the link layer
algorithm and rapid prototyping
library accelerated for RF
design
and rapid OAI CN • Code algorithm in Sionna
prototype • GPU accelerated simulation
• Multi-GPU support
OAI 5G L2+
to gNB SW Stack • Fast BLER charts → rapid iters
Aerial 5G L1 on algorithm

OTA gNB HW A100 GPU NIC CPU

testbed
and • Generate inference pipeline
Benchma • Integrate into Aerial L1 source
Fronthaul switch
rking
Remote radio unit
O-RU O-RU • Test and benchmark on ARC OTA
testbed

ARC

5G NR UEs
Implementation on ARC network

ARC: from simulation to bring-up in OTA testbed

 Research Community Extending ARC Stack
Northeastern University Integrating OSC RIC into the ARC Stack

• ARC researchers will

• Bring AI/ML and GPU acceleration to reality in L1
and L2 in real OTA ARC systems
• Extend the ARC SW stack

• NEU is our first example of the eco-system partner

extending the stack by integrating the OSC RIC and
returning this work to the ARC community

• Huge number of RIC uses proposed by industry and

academia
• Example: 20 uses-cases in O-RAN document [1]

Figure from from [2]

[1] O-RAN.WG1.Use-Cases-Detailed-Specification-v09.00
[2] M. Polese, et al, “Understanding O-RAN: Architecture, Interfaces, Algorithms,
Security, and Research Challenges ,” https://doi.org/10.48550/arXiv.2202.01032
 Classical Iterative detector-decoder (IDD) MIMO Rx
ETH and NVDIA; Implemented in Sionna; Simulation acceleration with multi-GPU DGX Station

Prototype and simulate in Sionna Classical iterative detector-decoder (IDD) MIMO receiver
Construct new IDD by stacking
Keras layers Iterative Detection and Decoding
LE D
Dec or LDec
Train on DGX station
ARC L1 CUDA source code
enabling → new IDD algorithm ..
. y ỹ LE
OFDM Noise Data Det Channel
can be coded and integrated in b
D
RF Rx Whitening Detector Decoder
OTA system
Focus on your innovation and the Cn0
complexity of your problem not e
all the rest of the stack that is Channel H
Figure from from [1]
required for OTA operation Estimator

[1] R. Wiesmayr, C. Dick, J. Hoydis, C. Studer, “Deep-Unfolded Interleaved Detection and

Decoding (DUIDD) Multiple-Input Multiple-Output (MIMO) Recievers,” IEEE Asilomar Signals,
Systems and Computers Conf. Asilomar, Oct. 30– Nov. 2, 2022. https://arxiv.org/abs/2212.07816

Develop algorithm in Sionna, OTA benchmark on ARC

 DUIDD: Deep UnfoldeD Interleaved Detection Decoding
ETH and Nvidia; Implemented in Sionna; Simulation acceleration with multi-GPU DGX Station

DUIDD employs deep- unfolding

Principled iterative algorithm is
unfolded and trainable parameters
introduced that optimizes soft
information exchanged between stages

Trainable parameters also introduced in

LDPC decoder
Protoyped in Sionna trained on DGX
Station for extremely rapid design
turns and BLER chart production
Sionna design phase completed with
Figure from from [1]
code returned in Sionna public repo
DUIDD ML approach provides
significant advantage over conventional [1] R. Wiesmayr, C. Dick, J. Hoydis, C. Studer, “Deep-Unfolded Interleaved Detection and
designs [1] Decoding (DUIDD) Multiple-Input Multiple-Output (MIMO) Recievers,” IEEE Asilomar Signals,
Systems and Computers Conf. Asilomar, Oct. 30– Nov. 2, 2022. https://arxiv.org/abs/2212.07816
Current phase of development is ARC
integration

NVIDIA workflow: from algorithm design in Sionna to OTA on ARC

 What can you do with Open-source L2?
Example: Multi-call scheduling

• ARC open source L2 is the segue to • Conventional solution: a single-cell scheduler per cell, making locally-optimized
developing and benchmarking new GPU resource allocation decisions
accelerated MAC layer algorithms in OTA
testbeds … … …
L2 L2 L2
Scheduler Scheduler Scheduler

L2
… Scheduler
L2 L2
• Example: Joint scheduling of multiple Scheduler Scheduler
coordinated cells: a multi-cell L2 L2
L2
scheduler making globally optimized Scheduler Scheduler
Scheduler
resource allocation decision
… … …
• Compute heavy: workload matched
to GPU rather than CPU
• Need access to L2 source so new
• Joint scheduling of multiple coordinated cells: a multi-cell scheduler
algorithm can be tightly coupled to
making globally optimized resource allocation decisions
rest of L2 … …
…

L2 Multi-
… Cell
Scheduler

… … …
S52156 A bridge to 6G - Aerial Research
And Innovation Platform

We look forward to hearing from developers,

researchers, and industry on plans to use,
extend and innovate using ARC. Please
reach out with your use cases
[email protected]

Useful Links

1. NVIDIA Aerial SDK

2. NVIDIA Aerial Research Cloud Docs
3. OAI website
4. Y2023 Spring MWC-Barcelona NVIDIA-OAI
Interop Demo
5. MWC-Barcelona OAI blog