Laurent Leboucher, Executive Vice President Networks & Group Chief Technology Officer at Orange, says that Orange is “taking seriously” the use of a new architecture to support AI-native RAN processing, exploring the potential to reduce costs and energy consumption and to increase spectrum efficiency.
Leboucher said that while focus on certain aspects of the AI-RAN model have created “confusion” and can be a bit misleading, “part of the story is very serious, and I take it very seriously.” Leboucher spoke to TMN during FutureNet World, held in London on 21-22 April, and a week after Orange’s announcement that it would be exploring AI-RAN architectures with Nokia and Nvidia.
Leboucher explained, “We’ve decided that we want to test new architectures, moving away from the traditional ASIC implementation and implementing the radio algorithm on general purpose chip sets. We can use those chipsets because they provide parallel processing so you can do channel estimation, scheduling, beamforming using the same algorithms, but implemented on the parallel processor.”
He added that this kind of processor could support a “completely different” way to implement the Layer One of the radio so that instead of using the traditional algorithm the radio is using “real AI”. But the first step is first step is more to look at how existing algorithms can be implemented on a new architecture
“Having a GPU kind of infrastructure processor could be relevant, but it doesn’t mean that it has to be a big GPU: it’s not a Blackwell implementation that you need. Some vendors, for instance, are saying you can do that on a CPU and to some extent, they’re right. So what we want is really to understand that – understand how we can also scale and if there is an opportunity to reduce the cost to optimise and reduce energy consumption, to increase also significantly spectral efficiency, which for us could be extremely relevant.
Leboucher said part of Orange’s exploration will be to understand the TCO of a new architecture – adding that in time he would like it to cost no more than the existing method.
“I’m sure that at the beginning it will be a bit higher. The question is it two times higher, five times higher, 10 times higher? So we need to understand that, we need to have the data point. I think, honestly, I think it’s a matter of maybe two, three years before it becomes something more production ready.”
We need to be extremely careful not to sell a dream here
As for further exploitation of an AI architecture, Leboucher is not opposed but still counsels caution. He maintains he distinction between exploring the benefits of general purpose processing and buying into the entirety of the AI-RAN hype, with its promise of sharing GPU capacity to support RAN and other, external, AI workloads. He points out that the operator can support sub 5ms latencies on an optical link between cities as far apart as Marseilles and Paris, perhaps negating the need for a very distributed AI architectures.
However, that’s not to say he doesn’t see any opportunities for integrating AI capabilities into the RAN.
“Looking at how we can combine those different algorithms and really push the line in order to increase spectral efficiency could potentially make the difference. And on top of that, looking maybe at other use cases that could benefit from this infrastructure – which is not always used at 100% and, by the way, never used at 100% – yes, why not? For me, at least, it’s not the first intent. My first intent is really radio. And when we look at other use cases, the question is for what?”
Answering himself, Leboucher posits that one of those “for-what’s” could be sensing?
“We need to be extremely careful not to sell a dream here but I think there is potentially an opportunity with sensing. The reason being that for radio sensing you need to combine detection with the radio parts. Your network needs to behave like a radar and at the same time your radar needs to be smart, so you need to be able to understand the pattern and react. If you take, for instance, our frequencies 3.5 gigahertz, and maybe later 6 GHz, if you want to detect small objects it will be not easy to make the distinction in between a small drone and a bird.
“So if you want to make that distinction, you need to add other mechanisms to use AI in order to make it work. So in that case moving all the images in real time, to the center, to the cloud, it would be nonsense. So having that close to the edge, maybe not at the far edge, but somewhere towards the edge, makes sense.”
Comments
0