People seem to think that AI-RAN is the same as saying “GPUs in the RAN." Let’s look a little closer at this…if we peel back one or two layers of the onion, we can see where GPUs fit and where they don’t fit.
AI-for-RAN
AI/ML techniques are already in commercial service, boosting capacity in the radio access network (RAN). Multiple RAN vendors have tested this technology, and Mobile Experts has been given access to see trial results.
We’ve been reporting for the past three years that 15-20% capacity boost is available from simple AI/ML algorithms running real-time in the RAN software. To be clear, this is running on custom silicon with CPU and DSP cores, not GPUs.
There may be a future for GPUs in centralized RAN configurations, for example in coordination of beams from multiple sites to optimize the RAN even more. But I suspect that these ideas can also be exploited with custom silicon or x86 options, because the RAN has a finite number of inputs and outputs. A human-language model has a nearly infinite number of possible outputs. A RAN model optimizes about 200 factors and is limited to very specific decisions.
AI edge computing
I have looked high and low. I can’t find any AI-inference customers that are begging for low latency in a public telecom network. So, my first reaction to “AI-and-RAN” is that it’s a solution looking for a problem.
The Nvidia dream is already fizzling, as the operators that they’re better off with regional racks of GPUs instead of GPUs at every site.
There are interesting nuances to this trend. One will be the rise of Physical AI. T-Mobile is partnered with companies like Serve Robotics. They’re the people that make the happy-face robots that deliver food along the sidewalks of major cities. This is a very interesting case, where the autonomous robot makes navigation decisions on-board. Everyone has seen deeply satisfying videos of these robots getting stuck in a hole, or struggling to cross a street.
These little robots will improve with more training over time. But unlike self-driving cars, these little delivery robots can’t afford to upgrade to 200+ TOPS of compute power. A mom-and-pop restaurant simply cannot pay $50K for a delivery robot. As a result, I am expecting a hybrid approach: the robot will drive happily along the sidewalk most of the time, but will send a few frames of video up to the network when it reaches an intersection, to get access to a much larger set of training data and a bigger analytic model.
This is a pretty typical Physical AI application that we can envision today: a little robot that runs autonomously, but needs help at times. It’s important to point out that in almost every case, the latency requirement here is NOT 10 milliseconds. A food delivery robot can wait a full second — or even five seconds — at a street corner to get a better decision on how to cross. So the ideal place to position the AI model would be in a regional data center, not at the tower. Heck, most of the lag time is compute latency, so adding 30 ms of network latency is not a big deal.
Over the next 10 years, physical AI automation will clearly get more sophisticated, and we will need a big improvement in dexterity. That will lead to more real-time decisions. I believe that the regional GPUs will migrate closer to customers when those applications become more clear, but that’s 10 years away.
There is one case where the GPUs are useful at the radio site today: On-premises enterprise scenarios. GPUs are already widely used for on-prem edge computing. Industrial customers use GPUs to control robots over wires today, and as they marry their robots with a private 5G network and a GPU, we will see on-prem networks that look a lot like mobile edge computing.
It’s very possible that these on-prem robots will end up walking out of the factory someday and performing other tasks in a more public environment but that will also take time.
In the meantime, the focus will be on custom silicon and x86 in the RAN, with private 5G networks connecting GPUs for enterprise robotics.
So, sorry Jensen, there’s no need for GPUs in the telco RAN.
Joe Madden is principal analyst at Mobile Experts, a network of market and technology experts that analyzes wireless markets.
Opinions from industry experts, analysts or our editorial staff do not represent the opinions of Fierce Network.