AI and 5G Are Better at the Edge

Deploying for proximity is essential to meet latency requirements and get the most value from both technologies simultaneously

Oleg Berzin
Kaladhar Voruganti
AI and 5G Are Better at the Edge

The IT world is filled with exciting technologies that promise to unlock new possibilities for enterprises. In some cases, the confluence of two emerging technologies amplifies the benefits of both technologies. AI and 5G are perfect examples of such complementary technologies. Each one has tremendous potential, but they’re even better when used together.

Why location matters for AI

The AI workflow involves ingesting large amounts of data from multiple sources, using that data to train models, and then using those models to produce automated, data-driven outcomes. Increasingly, as more data gets generated at the edge, the different steps (workloads) in the AI workflow are often performed in different locations based on performance, privacy, flexibility and cost requirements. This trend is known as distributed AI. Many businesses are now working with distributed AI orchestrators to help them move their AI training and inference workloads to the appropriate locations.

Model inference and model training—the two workloads that make up the iterative process shown in the workflow graphic above—have very different requirements from one another. To put it simply, model training is more resource-intensive, so it typically runs in a large data center or on the public cloud. In contrast, model inference is more latency-sensitive, so it typically runs at the digital edge, where it has closer proximity to data sources.

Why location matters for 5G

Like with AI, being successful with 5G depends on having the right infrastructure in the right places. The promise of 5G is all about enabling enterprise-class wireless services, whereas previous generations of wireless networks could only support consumer-grade voice and text services. When it comes to enabling these enterprise services, the user plane function (UPF) is among the most important components of the 5G network infrastructure. The UPF is responsible for de-encapsulating 5G user traffic so that it can move off the wireless network and on to outside networks such as the internet or cloud ecosystems.

Since the applications that 5G users want to access live on those outside networks, it’s essential to have reliable, low-latency connectivity where the UPF resides. For this reason, moving UPFs from their core networks to the digital edge is one of the most important steps that telco operators can take to unlock the full value of their 5G infrastructure.

5G helps AI move away from device and on-premises infrastructure

Many AI use cases have strict performance requirements; one way to meet these requirements is to perform inferencing on the device itself or using an on-premises server stored very close to the device. These kinds of servers are often found in the closets of stadiums, retail stores, airports, and anywhere else AI data needs to be processed quickly. This approach has its limitations: Doing complex AI inference processing on the device can drain the battery quickly, and the AI hardware on the device is often not powerful enough to do the required processing.

Furthermore, many AI use cases require data aggregated from multiple sources. In many cases, there won’t be enough memory/storage space on the device to host the diverse data sets. Similarly, doing AI inferencing in the on-premises closet has issues around physical security, physical space limitation, inability to provide the required power, and higher OPEX to maintain the hardware.

Since 5G networks provide high-bandwidth connectivity, it’s now possible to host AI inferencing infrastructure and also cache the required data sets at the 5G infrastructure located close to where the data is generated. Thus, AI inferencing tasks can be moved from device and on-premises locations to the 5G multiaccess edge compute (MEC) location at the nearby network service provider (NSP) 5G infrastructure in the same metro.

Being colocated with the 5G network helps satisfy the latency and bandwidth requirements of the application while also allowing enterprises to move their AI infrastructure away from the device or on-premises closet. Depending on the carrier 5G deployment architecture and application latency requirements, the 5G MEC infrastructure could be located in a micro data center (such as a cell tower), a cloud 5G Zone (such as AWS Wavelength) or a macro data center (such as an Equinix IBX®).

Use case: Data aggregation at the edge for smart parking

One example of a use case that requires data aggregation from many different sources is a smart parking scenario:

  • A low-power parking spot sensor can generate a message that a car has just parked.
  • A video camera can start streaming the video of the driver unloading shopping bags from the car.
  • After a video analytics tool determines that the driver is a resident of the building, the doors can open automatically.

In this example, the different data sets would each be carried by different networks (e.g., LoRa for the parking sensor, 5G for the video feed and WiFi for the smart doors). To enable the intelligent actions in the correct sequence, the AI inferencing would need to ingest data from all those different networks.

The edge is where many different networks intersect; it’s also where the traffic is initially aggregated and processed using edge compute applications. It’s therefore desirable to deploy AI inferencing in the same locations where edge compute resources are deployed, and where different networks intersect.

Furthermore, typical edge deployments are built using a hybrid architecture. This means that some processing happens on edge compute infrastructure and some in the core cloud. The same hybrid architecture is relevant to AI, where the core infrastructure used for model training is connected to the edge infrastructure used for inferencing. This enables a more dynamic link between the two, and therefore, more accurate outcomes.

AI enables better slicing and maintenance for 5G networks

One of the most powerful aspects of 5G technology is that it allows NSPs to perform network slicing, essentially offering different classes of network service for different classes of users and applications. Today’s NSPs can apply predictive analytics backed up by AI models to enable smarter network slicing than ever before.

To do this, NSPs can collect metadata on different applications, including how those applications perform under specific network conditions. When the 5G infrastructure and AI models are both located at the edge, it’s easy to get predictive insights about what quality of service (QoS) different applications might need, and classify them into different network slices accordingly.

In addition, NSPs can pull log and utilization data for the network and use it to train AI models that support proactive maintenance and management. These models can help detect conditions that indicate a possible service outage or surge in user traffic. The network can then automatically react to prevent the outage or provision additional capacity. Again, having both 5G and AI infrastructure at the digital edge is key to making the most of this capability.

Use case: AR/VR for predictive maintenance

Companies that perform maintenance on vehicle fleets are increasingly equipping their technicians with AR/VR goggles. These goggles allow the technicians to capture video and stream it directly into a video analytics tool. To get the best analytics results, AI processing needs to happen in real time. This means that there needs to be compute infrastructure deployed in close proximity to where the video is captured.

As previously mentioned, doing AI inferencing in the goggles themselves would not be practical. Storing servers in a closet would also not be ideal, for various reasons:

  • It can lead to higher OPEX costs if there are multiple maintenance locations in a metro, because the company will have to send technicians to maintain the hardware in each of the locations.
  • It limits flexibility, as there’s no easy way to scale up capacity as demand increases.
  • It creates physical security headaches, since the company has to protect the server themselves.
  • Finally, there’s the simple fact that most businesses would rather not have all their closet space taken up by servers.

5G offers a better way. A 5G-enabled MEC server can be placed further away from the maintenance facility while still providing high bandwidth and meeting very strict latency requirements. The maintenance company could deploy their 5G MEC servers at a colocation data center in the corresponding metro location to ensure they’re near the facility, without having to be in the facility. This can help them achieve all the benefits that on-premises hardware can’t: cost-efficiency, flexibility and physical security.

Use case: AI-assisted predictive workload relocation for vehicular services

Vehicular services—commonly referred to as vehicle-to-everything (V2X)—increasingly rely on ultra-reliable, low latency (URLL) communications between the moving vehicle and critical edge computing workloads. These workloads are responsible for delivering enhanced levels of autonomy, ranging from predictive insights (L3 – Adaptive and L4 – Semi-Autonomous) to fully intelligent operation (L5 – Autonomous) as envisioned by the automotive industry.

NSPs have architected and designed their 5G networks to support URLL communications for vehicular mobile services. If the application workloads are running in suboptimal edge locations (e.g., edge data centers), it may create excessive distance between those workloads and the moving vehicle. This distance may negatively impact the performance of critical application traffic between the vehicle control system and the application workload running on MEC resources.

AI models can help predict the edge computing (or MEC) location where the application workload and its context should be moved in order to meet latency, throughput and reliability requirements. This AI-assisted workload relocation decision can be based on data gathered from multiple systems, including:

  • 5G network telemetry (e.g., signal conditions, handover events, available QoS)
  • Vehicle telemetry data (e.g., location, speed, heading)
  • MEC telemetry data (e.g., location, capacity, reliability)
  • Metro packet/optical network telemetry data (e.g., route latency and bandwidth between MEC locations)

The outcome from the AI-assisted computation can identify the optimal MEC location to replicate the critical application workload in coordination with the vehicle movements, 5G network conditions and MEC resources. Conversely, if the outcome of this computation is negative (no possibility to meet the requirements), the vehicle’s autonomy level can be reduced.

Deploy at the edge with Platform Equinix

Platform Equinix® offers data centers in 70+ metros across six continents, so we make it easy to deploy AI inferencing and 5G infrastructure in all the edge locations that provide the best results for your 5G and AI workloads. We also offer digital as-a-service infrastructure that can help simplify and accelerate your 5G/AI rollout, including:

  • Equinix Metal® for single-tenant compute and storage capacity when and where you need it
  • Equinix Fabric® for on-demand virtual connections across your own distributed infrastructure and within your partner ecosystem
  • Network Edge for virtual network functions (VNFs) from top vendors, including 5G-enabled virtual routers

Finally, Equinix has the vibrant partner ecosystem needed to maximize the value of your 5G and AI deployments. For instance, we partner with 2,100+ NSPs across the globe. We’re intimately familiar with how to help those NSPs modernize their networks for the 5G era, and we also know how to pass the power of those 5G networks on to our enterprise customers.

In addition, we offer cloud-adjacent data centers that can provide proximity to all major cloud hyperscalers. This enables better results for distributed AI. Customers can do their AI training in the cloud, and subsequently move their models to an Equinix data center in the same metro. They can move data between the clouds and the edge locations using a private, secure Equinix Fabric connection.

AI and 5G are just two examples of how leading companies are using distributed, interconnected digital infrastructure to achieve transformational results. For a closer look at how they’re doing it, and how you could do the same, read the Leaders’ Guide to Digital Infrastructure today.

Avatar photo
Oleg Berzin Senior Distinguished Engineer, Technology and Architecture, Office of the CTO
Avatar photo
Kaladhar Voruganti Senior Business Technologist
Subscribe to the Equinix Blog