11 min read
While I was thinking of a longer post on the recently concluded IEEE GlobeCom 2025 in Taipei, I could not find time. But here it is, finally. Due to the limitations on the LinkedIn post size, I’m switching back to my blog. While my blog page has been there since 2009, most of my thoughts were shared in the shorter format on LinkedIn.
For those who don't know, IEEE Global Communications
Conference (GLOBECOM) is one of the IEEE
Communications Society’s two flagship conferences (the other is IEEE International Conference on
Communications (ICC)) dedicated to driving innovation in nearly every
aspect of communications. Each year, more than 3,000 scientific researchers and
their management submit proposals for program sessions to be held at the annual
conference. After extensive peer review, the best of the proposals are selected
for the conference program, which includes technical papers, tutorials,
workshops, and industry sessions designed specifically to advance technologies,
systems, and infrastructure that are continuing to reshape the world and
provide all users with access to an unprecedented spectrum of high-speed,
seamless, and cost-effective global telecommunications services.
This year's conference was themed “Sustainable
Communications for Ubiquitous Intelligence”.
As many already know, I was invited to speak at the parallel event - Open RAN summit, under the industry program (9th to 11th December), but I had the privilege of attending the workshops and the tutorials on 8th and 12th.
Since there were multiple parallel sessions (except the keynotes), I had to select the tracks. The following are the ones that I attended;
- Tutorial on Scalable 5G/6G ORAN, MIMO, and RIS innovations for commercial 5G/6G communications
- Tutorial on Beyond classical security:
Quantum Technologies for Trustworthy storage and communications
- Open RAN summit
- All the Keynotes
- Industry panel on NG optical networks: How to build a sustainable network in support of emerging requirements
- Tutorial on Foundation models for communications systems
- Tutorial on Unleashing GenAI: addressing scalability, security, and cyberinfrastructure challenges
Check https://globecom2025.ieee-globecom.org/program/program-glance for the full program.
Below are my key takeaways from the ones I attended. I list
them under different headings for easy reference. As you can see, I selected
the tracks to get the best coverage of all the current topics in communications
today. In this way, I was able to get a bit of everything. 😉
In my view, the whole conference was more towards 5G and
AI.
There was an exhibition area where operators, OEMs, vendors, and researchers demonstrated their products and capabilities.
5G/6G
One of the highlights in 5G/6G was the vendors that
manufacture equipment and testbeds needed for 5G/6G research. These include
companies like TMYTEK, Allbesmart, Emerson, NI, Keysight. One of the
highlights was the collaboration between these companies to come up with groundbreaking
platforms for testing and research. Some of these platforms included;
- 5G/6G FR3 OAI testbed
- World’s 1st 5G FR2 SA Open-Source Platform
- Hands-on 5G kit
- mmW-SDR solution
- mmWave Joint Communication and Sensing (JCAS/ISAC) Testbed
- 4 x 4 MIMO Testbed with multiple 28 GHz beamformers
This, I think, is a very vital part of this specific
ecosystem. And the other beauty was that almost all such platforms were based
on open-source software. These include;
|
Radio Access Network (RAN) |
Core |
User Equipment (UE) |
|
OpenRAN Gym, which implements
an Open Toolbox for Data Collection and Experimentation with AI in O-RAN, is another
good project.
With regard to the security of wireless power used in test environments,
it was noted that 10/20 dBm was safe enough.
The major challenge in designing wireless systems is
striking a balance between the spectral efficiency (bps/Hz/km^2), which is interference-limited,
and the energy efficiency (bits/J), which is noise-limited.
As we move to mm wave few challenges arise.
- High path loss and signal blockage
- Hardware design constraints
- Beam alignment and tracking
The importance of synchronization in mobile networks,
including Private 5G, for stressed by many.
6G starts from 3GPP
Release 21. 6G will be the 1st generation to be fully open,
intelligent, virtualized, and interoperable from day 1. 6G KPIs will be
quite different from those of 5G. The KPIs include;
- Communications
- AI and computation
- Localization and sensing
- Dependability (availability, reliability, safety, integrity, maintainability)
One rational for 6G is, current traffic growth (54% annual
growth of mobile traffic from 2020 to 2030), which cannot be addressed with the
present 5G. This is also closely related to the new Moore’s law (every
year the model size becomes 10X).
Reconfigurable Intelligent Surface (RIS)
RIS seems to be a good solution for indoor signal coverage
and also as a cost-effective solution to do away with a smaller number of
towers. The applicability of RIS is not only for Terrestrial Networks (TN)
but for Non-Terrestrial Networks (NTNs) as well, the surface being used
on LEO satellites. Metamaterial can be used to build RIS on different
structures, including buildings.
While traditional parabolic antenna surfaces can also be
used to direct a wireless signal to a different direction, RIS is much cheaper
than parabolic antennas.
Integrated Sensing And Communications (ISAC)
ISAC is a combination of localization and sensing
(conventionally done through Radio Detection and Ranging (RADAR) and Global
Navigation Satellite System (GNSS)) with communications. Historically,
these 2 fields were in different verticals, but with ISAC, they come together
for the 1st time. Navigation with TNs is a stronger alternative to
GNSS. Terrestrial wireless/mobile networks precisely determine the UE's
location and positioning. ISAC can provide accuracy up to 1 cm level.
In an ISAC, the Tx system/node transmits a frequency (f1) in
the forward direction (towards the UE and Target - downlink) with data
modulation. While the UE communicates back (uplink) with a different frequency
(f2) with the node for communication, the target reflects the same frequency
(f1) for sensing. In this way, we use the same frequency (f1) for both
communication and sensing. The reflected f1 can be received by multiple BTS via
back scattering.
The design of an ISAC system is all about joint waveform
design, one that is not a pure RADAR system (RADAR-centric design that uses
deterministic waveforms) or pure communications system (communications-centric
design that uses randomness). However, there are certain tradeoffs.
- Sensitivity
- Degrees of Freedom (DoF) – this is good for communications, but not good for sensing
- Spectral shape
- Randomness (according to Shannon, randomness is essential for communications, but this is not good for RADAR)
Wireless sensing can also be a solution for Non-Line of
Sight (NLoS) issues. Robotics uses Light Detection and Ranging (LiDAR)/camera
for Simultaneous Localization and Mapping (SLAM) to find the location.
Radio SLAM is an alternative at lower frequencies (compared to light at THz).
So, in the future (or even today), communication is more
than mere data transmission. It includes location and sensing as well. After
all, understanding the physical environment can enhance communication systems. ISAC
in the future might even replace surveillance cameras and sensors.
AI is also related to ISAC – AI to enable ISAC and ISAC
to enable AI.
Open RAN
My friend Dr. Chih Lin-I, CMCC
Chief Scientist of Wireless Technologies, China Mobile Research Institute, and
the TSC Co-chair of O-RAN alliance, delivered
the 1st keynote of the O-RAN Summit.
Open RAN is all about being open, virtualized,
intelligent, and interoperable. O-RAN alliance focuses on technical specs,
testing & certification, and the software community (with the Linux Foundation with Apache 2.0 license). It
is important to note the difference between paradigms of hardware and software.
|
Software |
Hardware |
|
Information Technology (IT) |
Communications Technology (CT) |
|
Open Source |
Standard Development Organizations (SDOs) |
|
Function first |
Performance first |
|
Fast prototyping (CI/CD) |
Slow multi-year standard development |
|
De-facto standards |
Official de facto standard |
While 4G got the open-source software platforms years after
deployment, and 5G got them nearly 5
years after deployment, 6G seems to be concurrently using open-source software
platforms and AI/ML.
The open-source projects like ONAP,
k8s, Sylva,
Duranta, OCUDU, etc, were highlighted here. However, one of the challenges is
that the Mobile Network Operators (MNOs) are not yet fully convinced of
the business value of open-source.
There are over 900 O-RAN specs to date. These span
across the full e2e Open RAN environment: O-Cloud (the foundation hosting
environment for all open RAN virtual/software components), O-RU, open
fronthall, O-DU, O-CU-CP, O-CU-UP, Near Real Time RAN Intelligent Controller
(RIC) (near RT applications (xApps) are here), Non RT RIC (Non RT applications
(rApps) are here), Service Management and Orchestration (SMO). The SMO
will become a G-agnostic intelligent automation hub.
Source: https://www.o-ran.org/about
O-RAN alliance defines the interface differently from 3GPP.
While 3gpp go as N1, N2,…, O-RAN goes as A1, D2, E1, E2, F1, O1, O2, …).
The current O-RAN deployments include the MNOs like NTT docomo (50 M subscribers), AT&T (with Rakuten), TELUS (they will be 50% O-RAN by 2027 and
100% by 2029 with ORIC), and Bell. T-Mobile will soon release an RFQ. Both
NTT docomo and AT&T use open fronthall. Check the O-RAN map for the updated details.
The open RAN hardware vendors play a vital role, too. Specifically
for O-RU, vendors like Benetel and Foxconn (Hon Hai Technology group),
MYTEK, Pegarton 5G are significant.
I met most of them in the exhibition area.
Satellite
The part satellite is going to play in future communications
was highlighted by many. So, it’s going to be a combination of terrestrial and
NT like High Altitude Platform Systems (HAPS ) and the LEOs at around
600 km; all delivering 5/6G services.
There were some ideas on Satellite Computing Networks (SCN).
The satellites collect large data sets, such as those related to agriculture,
disaster, and weather. SCN could be the next battlefield of AI. However, the
following limitations exist
- Privacy, security
- Data communication bottlenecks
- Hardware limitations
It was also noted that Starlink
by Sep 2025 had 7.1 M subscribers in 150 countries. Today they have 9 M.
AI
The sessions on AI were, in a way, good for refreshing my
knowledge and also gathering new knowledge.
·
ML
o
Supervised learning
§
Classification (assignment)
§
Regression (prediction)
o
Unsupervised learning
o
Reinforcement learning (RL)
On the side of AI, it was noted that 1/3rd all traffic by
2033 is going to be AI-related. AI said to generate ~1000 EB/month by then. The
analogy given was: if 1 GB is 1l of water, this traffic is like the Pacific
Ocean.
The Internet might become a collection of AI agents in the
future. IP packet communication might become token communication.
Moving from Agentic AI to AI agents is like
moving from the executor to the thinker. AI Agents are for people, enterprises,
and robots (physical labor). Agentic memory systems are of different forms: short-term
memory on UE and long-term memory on the cloud.
Inferring in space was also discussed, but then, how to get
the results back to Earth was a problem unsolved.
The energy required to run AI is getting higher and higher
with new GPUs. Ex:- Nvidia H100 at 60% utilization needs 3.68 MWh/year.
AI-RAN alliance, which
talks about AI for RAN (Improve the RAN), AI on RAN (supporting
AI applications), and AI & RAN (convergence), was also in the
discussions. It’s the integration of AI into RAN to boost performance, improve efficiency,
and enable new intelligent services.
vRAN -> Open RAN -> AI-RAN
A new term – Business Grade AI was discussed, which
needs to have security, explainability, and trustworthiness.
The importance of the Model Context Protocol (MCP) to
communicate with an Agent and the Agent-to-Agent (A2A) protocol to
communicate between agents was considered important.
Following Agent networking challenges exist.
- Token capacity (communication and computing tools)
- Wall-clock time (latency)
- Trajectory complexity (reasoning and execution complexity)
The running location of the foundational models (ex:-LLMs)
was discussed. The choices were Cloud and Edge.
AI-nativeness? architecting the network with AI as a
core component from the ground up, embedding data pipelines and automation.
AI in the sky was another idea presented. The rationale
is the abundance of free power. The network proposed is Free Space Optics
(FSO). However, cooling and solar radiation are a challenge.
Foundational Models (FMs) for communication systems
were also discussed. While in task-specific models, multiple models are
required for different tasks, in a foundational model, a unique model can be
used for many tasks. The idea is to have one foundational model for telecoms,
combining the existing Large Telecom Models (LTMs) and Large Wireless
Models (LWM), for example. The tech giants like Open AI, Google
provide foundational models (ChatGPT, Gemini). While every LLM is an FM,
every FM is not an LLM.
The idea of machine unlearning was also discussed.
The fact that GenAI doesn’t know how to forget is seen as an issue. Further,
GenAI doesn’t know that it doesn’t know.
The fact that the hyperscalers do not own data, although
they own the DCs and power, was also highlighted.
Like cloud auditors, the requirement of AI auditors was also
discussed.
Quantum communication networks
Quantum communications are not new. However, like quantum computers, it is not standardized yet and has no layers like the ISO/OSI.
For Quantum bit (qubit) communications, we need quantum
nodes and quantum links. A quantum link (photon/optical) will link
the quantum memories on quantum nodes. Therefore, quantum traffic is very
sensitive. The initial links will be less than 150/200 km without optical
repeaters. Otherwise, quantum error correction with repeaters/amplifiers will
be needed. While Shannon’s
capacity theorem is still applicable for quantum communications, the post-Shannon
information theory, which extends Shannon's foundational work by shifting
focus from what message was sent (content-based) to whether a specific message
arrived or a goal was achieved (identification/semantic-based), is more important
here.
Quantum memory is all about physics. They are implemented on
nanostructures like color centers in diamonds.
Quantum Key Distribution (QKD) is a critical piece of
quantum communications. The usage of a satellite for this is under discussion.
It looks like we will have the quantum internet by 2039 - the
time of 7G. The work of the Quantum
Internet Research Group (QIRG) of the IETF
is important in this regard.
- 5G connect things
- 6G connects intelligence
- 7G is for agentic/token communication
Optical communications
The physical limits of communications, especially the
Shannon’s and Moor’s were discussed.
The work of the Integrated Optical
and Wireless Networks (IOWN) global forum,
with regard to optical communications, was highlighted in the Next Generation
(NG) optical networks tutorial.
GNPy is
an open-source, wonderful optical design tool.
A lot was discussed about All Photonics Networks (APNs).
The rationale was as follows;
- · Large-scale DCs (concentrated energy)
- · Distributed computing (distributed AI DC architecture) -> avoid concentrated electrical consumption and leverage local renewable energy -> requires a good optical transport network
This makes the current trend of DCs moving where the power
is to DCs moving where the network is.
The real-life example of CHT (Chunghwa Telecom) in
Taiwan and NTT in Japan doing a
co-performance 4K video over an IOWN APN was remarkable. The performance is an
example of cyber (virtual)-physical systems.
On the side of AI, Remote Direct Memory Access (RDMA)
over APN might likely replace TCP in the future.
That’s all for now. Hope to see you again soon. As always,
comments are welcome.
No comments:
Post a Comment