Ray Le Maistre, TelecomTV (00:14):
Artificial intelligence is set to impact every telecom infrastructure domain and the radio access network is no exception with AI RAN one of the hottest tech development topics in the cellular infrastructure sector right now. I'm talking today with Vihang Kamble, CTO at Rakuten Symphony's RAN business unit to find out more. Vihang, great to see you today. Thanks very much for joining us. So tell me, how do you think AI RAN is getting incorporated into 5G?

Vihang Kamble, Rakuten Symphony (00:47):
Thanks, Ray. Thanks for having me here. And AI RAN in 5G is no more experimental. It's already delivering real value through some of the pragmatic approaches. And some of it, thanks to O-RAN Alliance because they could foresee that AI is going to be integral part of the telecom networks. And RAN intelligence controller base architecture has become a foundation for embed intelligence into the RAN network. And what we have seen is using some of the rApps for non-real-time AI RAN, that has become the framework for solving problems like mobility energy saving. As part of Rakuten Mobile, we have integrated some of the third party rApps. And what we have demonstrated is we have shown 25% energy savings by using our app on RIC platform. Now for the real-time AI RAN, which is running on the edge compute, now the debate between CPUs and GPU is opening.

(01:52):
And we understand that GPUs do have parallel processing capabilities and suitability for AI workload, but we have to be pragmatic because some of the 5G products out there are mature and doing those deployments once again on GPU may not be cost effective. So the approach and the strategy which we have taken is GPUs are essential for AI training, no doubt about it, but some of the new agent servers have very good inference capabilities. It has this advanced matrix multiplication ability. It has ability to do the floating point optimisations and inference engines. So we have used this capability to showcase some of the RAN problems and we took two use cases. One use case was detecting the uplink signal more robustly and that resulted in much lesser errors for the uplink signals. And we have shown that a voice quality has gone from really poor to very good using some of the AI RAN optimisations.

(03:04):
Similarly, for adaptive modulation and coding, we use AI scheme and using that we were able to show that error rate of 20% we were able to bring down to 10% and that's significant because that can reduce the throughput fluctuation user can experience. So we have already demonstrated that AI RAN could be implemented on CPUs with inference timeline of 60 to 80 microseconds. And that's going to be our strategy that for 5G and 5G advance use GPUs for training, but for inference use the CPUs, which are upping their game already in the system.

Ray Le Maistre, TelecomTV (03:46):
Okay. Interesting. So obviously that's something that's pertinent right now to mobile operators around the world, but they're also currently thinking about their 6G strategies. Do you think AI RAN in 6G will be much different from the kind of scenario you just discussed there?

Vihang Kamble, Rakuten Symphony (04:04):
Yeah, interesting question. Interesting question. And I think the real question is not whether 6G will use AI, but it's like how deeply the integration of 6G and AI RAN is going to be. And what we have seen in some of the standards forum, some of the operators are pushing for 6G as software only option, right? They want to use the current hardware for 6G, which is good it will save the cost, but it will delay the integration of 6G and AI RAN. So what we propose is 3GPP and O-RAN Alliance have to take some bold stands and not stick to the incumbents view. So about the protocol, there is a good scope that AI RAN could be part of the air interface protocol itself and that could help us optimise mobility. It could help us take care of some of the SON problems in the RAN. That's about the protocol.

(05:08):
It could make the current AI assisted RAN into AI native RAN. About the architecture, Rakuten Symphony had taken a stand that going forward there will be a separation of software and hardware and third party RUs could be used based on O-RAN Alliance option 7.2 split. We also believe that the baseband software will be run on COTS hardware. Now that architecture philosophy is becoming real in 6G. Essentially the debate between CPU and GPU, it means that the hardware is going to be COTS. So no matter whether CPU or GPU wins, our strategy is more or less reinforced in 6G. And that's where we think some of the differences will be that 6G will be very interesting, will be very close coordination between AI and 6G going forward.

Ray Le Maistre, TelecomTV (06:07):
Okay. Interesting. And obviously there's a lot of debate around those kind of developments right now as the 6G standard gets discussed and developed. On a broader scale, what is the radio access network of the future going to look like due to the introduction of 6G and the use of multiple AI RAN applications? Do we have a sense yet of what this future RAN infrastructure will look like?

Vihang Kamble, Rakuten Symphony (06:40):
Certainly. I think the 6G networks in our view is going to be autonomous, more software driven and we think that AI and 6G will evolve together. So AI native RAN where intelligence is built into the baseband software air interface and every node will be potentially will have some AI scope. So that's about the air interface. Now about the SMO, which is the service management and orchestration, we believe that some of the proposals or some of the baselines which are happening in 6G are going to be very critical. So service-based management architecture will make the new services launching very easy as well as very scalable. We also believe that intent-based networking could become the future of the network, which means that you could just say that please design or please configure a network with certain quality of service for certain network slice, and that's it. The network goes and configures itself.

(07:50):
Currently, that job takes tuning of hundreds of parameters, which needs a lot of manual intervention. So some of that will go away. So automation is going to be the main part of 6G networks and it will reach maturity levels. So if you know that we already have reached level four automation for some of the energy optimisations which we have developed in Rakuten Symphony, we believe that in 6G it will further move to level five optimisations.

Ray Le Maistre, TelecomTV (08:23):
Okay. That's an interesting development and a scenario that a lot of operators are thinking about at the moment. And another thing they're thinking about is how their radio access network infrastructure might be used to support multiple workloads. What types of AI workloads will need to run and be co-located at the RAN edge in the 6G era and what kind of platform approach is suitable for the provision of such services?

Vihang Kamble, Rakuten Symphony (08:57):
Yes. I think we believe that some of the AI workloads which could share the compute power with RAN will be the ones which need low latency and some kind of traffic management is needed. So one example could be ISAC, which is the sensing application, which 3GPP is talking about. Another could be XR or the augmented reality real-time gaming. Some of these have really good scope to coexist with the RAN compute node. And what we believe is the preferred platform going forward is going to be some kind of a hybrid architecture. CPU first deterministic platform for DU as well as RAN Edge, but GPU augmented edge clusters when needed could easily be added. So this allows for coexistence of this hybrid approach where CPU and GPU could both be used in optimal fashion in 6G networks.

Ray Le Maistre, TelecomTV (09:56):
Okay. So a lot of really important RAN infrastructure developments there that operators are thinking about as they evolve their AI strategies and look forward to 6G. So Vihang, thanks very much for your insights today. It's been great to talk to you and we look forward to seeing you again on TelecomTV in the future. Thank you very much.

Vihang Kamble, Rakuten Symphony (10:17):
Thanks, Ray.