AI and its older sibling, machine learning, have been around and in development in telecom networks for many years now. This is particularly true among the tools of network testing, measurement, monitoring and assurance, which have traditionally been the means of viewing the network: The foundations of characterization and validation, fault detection and root cause analysis. These have been painstakingly built to cope with the complexity of telecom, have evolved with every new spectrum band and G, all the ever-increasing acceleration of speed and data traffic. But the advent of generative AI and associated increased computing power has sparked new interest in how AI might be put to work within the telecom network. Operators expect to see benefits that include expanded automation across network planning and operations, as well as new paths to monetization.
RCR Wireless News reached out to Keysight Technologies for its perspective on the impact of AI in the network testing and assurance realm. The company has been integrating AI across its portfolio as well as supporting industry efforts such as the use of AI in open Radio Access Networks. The following Q&A was conducted with Joel Conover, senior director at Keysight, via email and has been lightly edited.
RCR: AI and ML aren’t new to the networking testing space. For some context, how has Keysight been working on the development and application of AI in network testing in the past few years? What have you seen change in this space in terms of technology, or interest, in the last 12-18 months?
Conover: AI applications are driving an entirely new set of requirements in our customers’ network equipment and in their network architectures. The interest is driven by massive investments in AI data centers build for training AI models and operating AI models (inference), and a desire to maximize the return on investment in those data centers, GPUs, and associated infrastructure. We’ve partnered closely with some of our largest customers to develop new emulators and measurement techniques to accurately emulate the unique parameters of AI data center workloads, and to measure overall system performance. A key challenge is developing metrics that help pinpoint what part of the system is limiting performance, so that the networking, compute, storage, and higher level applications and protocols can all be optimized to maximize GPU utilization for both AI training and AI inference operations.
RCR: Can, or is, generative AI, being applied in network testing? Could you give examples of what that looks like/use cases?
Conover: Our focus today is on enabling customers to build, train, and operate their AI models faster, and gen AI is at the forefront of the models that our customers are building. The use case for network testing is emulating the unique properties of that environment, and delivering it at a scale we’ve never seen before.
First, we had to come up with new models (algorithms) that emulate the I/O behavior of AI workloads. This can include protocol interactions between clients/hosts, GPU processors, the network, and the algorithms controlling the AI training or AI inference processes.
Then, what made this so interesting was the scale needed. Customers need to emulate anywhere from hundreds to hundreds-of-thousands of GPUs, which could span one rack or hundreds of racks of servers, and correspondingly large networks. AI data centers are already pushing the limits of even the fastest commercial ethernet technology available today – 800 Gigabit Ethernet. This will be a driver for 1.6T Ethernet and beyond. The effort is about a lot more than bandwidth, but that’s in part because the system can use more bandwidth than is possible with today’s technology – meaning bandwidth is a more constrained than ever.
Putting together a solution capable of emulating a system that complex, at that scale, is something no-one has ever tried to do. But we’ve set out to do it.
RCR: What do you see as the biggest risks or challenges to broader, deeper use of AI in telecom test and measurement?
Conover: AI is being infused into many aspects of communications technology – it shows particular promise in predicting channel conditions, essentially creating new forms of “smart radios” that can achieve higher throughput and/or longer distances by incorporating machine learning in the radio itself. Developing tools to simulate and tune those models early in the R&D cycle is already underway, paving the road for developing 5G advanced and 6G radio technology. Ethernet networks are also being infused with AI to improve route selection, manage link congestion, and more. As a design, test, and emulation company, we have to be looking at all phases of the product lifecycle, from those early simulations all the way through network operations, and develop techniques to fully exercise and even train AI models in the R&D lifecycle.
RCR: AI has two aspects within test: Both the use of AI in test tools, as well as test tools being used to validate and monitor AI. What opportunities does Keysight see in both of those areas?
Conover: Our unique breadth in design, emulation, and test solutions creates lots of opportunities to incorporate AI in our tools, as well as using the tools to test, validate, and monitor systems infused with AI. We’re been using machine learning algorithms in our design suite for several years, for example. The latest versions are helping customers build better radios using machine learning to simulate and optimize impedance matching circuitry based on active load pull measurements. In measurement and analysis, we have machine learning suites built in our analytics offers. These ingest large amounts of data and can then identify anomalies in long-sequence captures, which help developer find elusive glitches in systems, for example. And GenAI can benefit all sorts of applications – from enabling engineers to use voice prompts to craft instrument automation code, to returning faster, more relevant results when accessing product support – as we’ve recently done on our customer support website. We’re just scratching the surface of all the ways AI can be infused in products to drive greater efficiency or insight.
