Also see: Top Edge Computing Companies
I give this question a definitive no. One could run down a game of semantic parsing, attempting to determine exactly what is meant by edge and cloud. But I think at this point we can broadly group the elements of interconnected nodes into endpoints, edge devices, and cloud components — with wavy borders between them.
That having been said, of what will the burgeoning edge consist? The best recent example I can think of is the edge video processor for home and commercial security, pet detection, facial recognition, retail analytics, fitness, and smart conferencing. This device, a discrete, stand-alone accelerator for inferencing applications is not an endpoint. It hosts endpoints, in particular cameras.
In the case of video processing, the smarts are needed to determine what’s in the image. So, surveillance cams don’t burp at everything, just the image types they care about, like, say, people. The cams could do this work, but shipping images to a nearby node with more storage, greater processing power, and larger artificial intelligence (AI) models allows one edge node to manage the input of 24 high-definition cameras.
Then, when the rare image that warrants closer attention is picked out by the edge image processor, it can be passed along to a centralized cloud component that has input from multiple sites, even more powerful models, and the ability to pass intelligence back to the edge nodes.
Evolution at the Edge
Thus, the role of the edge node is just beginning to evolve now. This space has been hotly contested by the main players in the silicon industry. Qualcomm, Nvidia, and Intel all want a piece of this market.
And they’re not the only ones, but the nature of this contest is related to these companies’ initial positions in the game. Through its mobile phone processor business, Qualcomm has been strong in endpoints.
Intel, which mostly owns PCs (which are traditionally endpoints, but can serve as edge devices), achieved its position in the cloud as the king of servers. Nvidia, traditionally in the graphics business, has added a substantial new cloud business selling banks of graphics processors to cloud customers for specialized workloads.
If you ask the question of where the intelligence should be, you can answer it easily: everywhere. We can assume all nodes will get smarter. That means the cameras themselves, the edge gatherer, and the cloud component. All will get smarter.
For years already, cloud providers like Amazon, Microsoft, Google, and Netflix have moved information to the edge of their clouds to make popular content more accessible. When Casablanca, the classic movie starring Humphrey Bogart and Ingrid Bergman, is suddenly popular again because someone famous wrote or tweeted about it, Amazon can move more copies from its core cloud to edge servers close to geographic markets where this sudden popularity is arising.
Where Does Processing Occur?
There is a principle that guides where to execute an analysis model. Essentially, processing should take place as close to the data source as possible for three good reasons: privacy, latency, and efficient use of network resources.
Cameras that decide what is a human and what is a pet generate images. If those images, which may be too big to be processed on the cam, are analyzed in a home edge device, the potentially sensitive pictures never have to leave the home.
Such analysis can be time-critical. If someone monitoring a house remotely from a security app needs to know whether a caller at the door is a delivery person or a burglar, there may not be time to send the image to the cloud for analysis. And then there’s the cost of moving large data files all over the place. Best if that path is as short as possible.
In the AI world, the two big tasks are training and inference. In training, the model is taught what is a human and what is a pet through a massive ingestion of correctly tagged images. After a certain amount of training, the model can pick out the one from the other; that is, a trained model can make correct inferences. Training takes a ton of resources and it most appropriately done in the cloud, which is far less resource-constrained than endpoints and even most edge devices.
But inferencing must be done on either endpoints or edge devices. On-device inferencing can be used for things like voice activation, text recognition, face detection, voice recognition, computational photography, and object classification.
But since AI models need to continuously evolve and improve to eliminate both false positives and false negatives, a corrective cycle must necessarily involve all levels of computing. In a federated learning model, the cloud-based aggregated model takes input from all downstream devices, improves its ability to correctly identify the object in question, and updates all the downstream inferencing models. Thus, the model is improved globally from more diverse data.
And edge devices and endpoints can do local improvement based on the specific data set at that location, which may differ from the global set that was used for the original training.
Bottom Line: Edge and Cloud Must Cooperate
AI is just one domain that illustrates the way all levels of computing — endpoint, edge, and cloud — need to cooperate for the best outcome. There are many more where the division of labor among computing elements makes broad sense: intensive and large-scale computing in the cloud, the offloading of local tasks or positioning of cloud copies at the edge, and fast, efficient computing on the endpoint.