Chip and system designers scramble to leverage existing and future standards as edge AI increases demand for faster data movement and greater reliability.
Key Takeaways:
The accelerating build-out of edge AI is starting to redefine how people interact with AI, shifting the focus from massive global data mining and analysis in huge AI data centers to faster results, greater efficiency, and much more targeted workloads at the edge.
In both cases, the emphasis is still on processing and moving data at blazingly fast speeds. But at the edge, there is less data to process, and the distances that data has to travel are shorter. Hyperscalers emphasize contextual search, massive simulations, and training of large language models. At the edge, goal may be as limited as feeding commands to a robot about how much pressure is needed to pick up an object, or telling a car to jam on the brakes because a pedestrian just darted across the street. Small language models that are domain- and workload-specific replace more generalized capabilities in LLMs.
There is demand for both, but as the edge takes shape, it is beginning to look very different from what OpenAI or Anthropic does. “It’s definitely not the AI of the hyperscalers,” said Ananda Roy, senior product line manager at Synaptics. “This is AGI (artificial general intelligence), but it’s also AI at ultra-low power levels. It used to be called tiny machine learning until a few years ago — ultra-low power with low memory requirements. So you can run small AI models such as predictive maintenance, or Wi-Fi sensing like detecting your environment, or bouncing signals off of a person to determine whether they’re present or moving.”
Wi-Fi plays an increasingly critical role in the edge’s evolution, a far cry from its original role for sharing internet access in the home or coffee shops. As more devices are added to the edge, it is by far the most popular technology for sharing access. And with Wi-Fi 7 and the forthcoming Wi-Fi 8 (expected late 2028), it is becoming more predictable, reliable, and secure.
“Wi-Fi was originally designed as a best-effort technology,” said Sivaram Trikutam, senior vice president for Infineon‘s wireless product line. “This is different than cellular technology, which was designed for certain levels of reliability. But because of where it’s being used today, like industrial robotics and automation of various sorts, there’s a need for very high-reliability Wi-Fi. It’s a utility now, and you cannot have blips.”
Despite the best efforts of cellular providers, Wi-Fi has pretty much won the battle for on-premises data movement. For years, cellular providers pitched 5G and 6G as the future of high-speed communications everywhere, forcing chipmakers to support both. In theory, 5G millimeter wave technology can reach speeds of more than 10 gigabits per second — basically, the speed of fiber — but reality is somewhat different. Millimeter wave signals attenuate rapidly, and unlike 4G LTE, mmWave signals do not go around corners or through windows, and they can be disrupted by moving objects such as people, trucks, or even thermal changes. The only solution to that problem, so far, has been to put repeaters everywhere and set up small cells. But the amount of infrastructure required for what is essentially direct line-of-sight communication, and the challenges of maintaining that equipment in sometimes harsh climates, sharply limits the usefulness of 5G/6G mmWave.
Wi-Fi also has the capability to stream data back and forth to multiple devices at the same time, and chipmakers are working on ways to speed that up even faster. This is especially important as the number of electronic devices connected to local Wi-Fi networks balloons.
“When you move from one spot to another that is covered by a different access node in an enterprise or a factory, especially for robots or other applications that require reliable connectivity, Wi-Fi is very beneficial because it can serve many different nodes that are connected to the same centralized point,” said Sassan Ahmadi, product manager at Keysight Technologies. “And now, with AI chips, you want to put those at the edge because that’s where you can gather the analytics and process them and make improvements in the network for traffic handling, optimization, mobility enhancement, and all those things where you require intelligence.”
Security
At a minimum, Wi-Fi improves security by allowing users to determine whether processing is done locally or in the cloud. That minimizes data leakage, and it makes it harder to steal data without physical access to it.
“If you are a defense company, you don’t want your data to leave your premises,” said Keysight’s Ahmadi. “Your edge is within your premises, or within your firewall or your security wall. You probably have some centralization, but you don’t want your data to leave. For the cellular networks, the edge is really the operator field office. One of the reasons you are putting some functionality at the edge, and not in the cloud, is because of the turnaround time to send and fetch data from the cloud to support time-sensitive services. You need the edge to be as close as possible.”
Additional security features is being built into Wi-Fi chips, as well. “The requirements from the industry are becoming more and more stringent regarding what security should be supported,” said Shishir Gupta, vice president of product marketing and customer engineering at Synaptics. “We define security from the ground up, from the hardware to the software. So we have a hardware root of trust, and we will have PSA (Platform Security Architecture) level 3, Arm Trust Zone and memory protection, and secure boot.”
Screenshot
Fig. 1: Synaptics’ Veros Wi-Fi 7 processor with Bluetooth. Source: Synaptics
Consistent performance
In its early years, Wi-Fi was never considered a critical technology. If an Internet connection froze, or there was a long lag time in searching the internet or updating an app, it rarely caused any problems. But as Wi-Fi takes on a more strategic role, those blips in service are becoming unacceptable.
“This is basically determinism,” said Infineon’s Trikutam. “It’s an outer bound for latency. I need to be able to say, with certainty, that ‘traffic of this nature takes no more than this much time to get to where it’s going.’ A best-effort approach, which is where Wi-Fi started, is that the data will eventually get there. If you go back 20 years and you clicked on Napster, you’d eventually get it, but you didn’t know if it was going to take 2 minutes for a song to download or 20 minutes. That’s what best effort means. It does its best, and you’ll get it at some point. From there, we’ve made a lot of improvements, and the name of the game over the last few years has been improving reliability rather than speed. The next revolution from here is going to be improving determinism. That’s what’s coming in Wi-Fi 8.”
That determinism applies to an increasing number of devices that are connected in the home, office, or some industrial setting. “I have 45 or 50 devices in my home,” Trikutam said. “All those devices have to work all the time without disconnecting. And if for some reason they disconnect, I expect them to automatically reconnect. That has to be seamless.”
Making that all work is more complicated than it may appear. Reliability, or determinism, is just one of several key pieces to enabling edge AI everywhere, and all of them are in a constant state of improvement. Another piece involves the physical medium for moving data, which can include everything from point-to-point microwave to copper to fiber optics, the latter of which is proving to be particularly useful for moving large amounts of data quickly both in the data center and at the edge.
“The edge nodes are connected to field nodes through fiber optics, typically a 10 gigabit/second or 25 gigabit/second fiber optic,” said Keysight’s Ahmadi. “Some operators still have microwave links, but many operators have moved to fiber optics, especially in China, Japan, and Korea. So there is enough capacity to move a large volume of data. If you don’t have that, or you have limitations in supporting that kind of capacity, that limits your edge. Instead of having, for example, 100 access points connected to one central node, now you may have only 20 access points connected to one central node. That also affects your ability to synchronize data, which is important as the user moves from point A to point B.”
Changing chip architectures
This opens up a whole new way of looking at semiconductor architectures, where the emphasis is on consistent and high-performance, moving data ever-closer to the processing elements, and adding inferencing capabilities to orchestrate and oversee all of the various pieces operating in real-time.
“There can be a very rich set of peripherals,” said Synaptics’ Roy. “It can interface with different kinds of sensors — light, temperature, and whatnot — and extract the relevant data and feed it into a machine learning model. And it can do something intelligent based on the inputs it gets, which was not possible earlier.”
The key here is optimizing one centralized source of data and all of the different devices that feed off that data, which is a non-trivial challenge. Synaptics’ Wi-Fi chip, for example, has three types of memory. “We have dedicated memory for connectivity RAM and connectivity ROM, which can only be used by the connectivity space,” said Roy. “So the MAC and baseband processing, and some of the stack will reside there, which is not open to the application. But then there’s the application SRAM, where you can load the data onto whatever application you are running. And then you will have execute-in-place from flash. So there will be external flash, where the main code memory will reside, and you will boot securely from the flash with minimum latency, and then execute from there without having to bring all the software from flash onto this chip. It can directly run off of the flash. This is distributed memory, so it makes it very efficient, both in terms of latency and power.”
On top of this, Wi-Fi 7 and 8 chips also need to support Bluetooth LE 6.0 (short-range communication), Thread (LP mesh network), and Zigbee (mesh network). Steering this data through this electronic plumbing maze at exactly the right time is a daunting challenge. As with chips developed for the cloud, those developed for the edge are becoming more specialized as new options emerge.
“Right now we have a couple of use cases in Wi-Fi sensing, Bluetooth channel sounding, which is like estimating distance,” said Synaptics’ Gupta. “Channel sounding essentially reads the phase and the tones of Bluetooth waves to figure out how near or how far another Bluetooth device is, with about 20 to 30 centimeter accuracy. The thought is that, if your device already has Bluetooth, why not put something like ultra-wideband or radar, which consumes more power, on there instead of adding the cost of another chip? There are lots of applications that only need that kind of accuracy.”
Conclusion
Wi-Fi is coming of age, but it’s not necessarily because anyone planned it that way. When Wi-Fi was first developed, AI was still the stuff of science fiction. The rollout of ChatGPT in November 2011 changed perceptions across a wide swath of technologies and business possibilities. AI at the edge is just beginning to gain traction, and with that comes an understanding of possible future directions and what they could enable, and what is required to get there.
Wi-Fi is a key element in this transition. “Edge is like the demarcation point where you guarantee round-trip time of data, which ideally should be less than 1 millisecond,” said Keysight’s Ahmadi. “For every kilometer of optical fiber, you have a 3-second delay, so if it’s 20 kilometers, then you have a 60 millisecond delay. Beyond that, you cannot guarantee anything.”
Related Articles
6G Rollout Will Be A Patchwork At First
Spectrum allocation, infrastructure development, and varying use cases will affect when and where this technology rolls out.
6G Line-Of-Sight Repeaters, Dots, And Reflections
Private networks and city-wide simulations show how 6G providers will reach edge devices via repeaters, small cell radios, and reflections.
New Antennas And Advanced ICs Needed For 6G
Higher frequencies bring more complexity to every part of the network, from tiny antennas to powerful base station processors.
 |
 |
 |
 |
 |
 |
 |
 |
Leave a Reply