Good, at least I got that right. Well, good to see you again. Embedded World, very interesting show for us. This is not a show that we have typically participated in, and Industrial is an area that we've started to focus on quite seriously over the last year or so. We did a couple of things over the last six months. First, we reorganized our entire product portfolio, made it very focused on a variety of different processors: consumer, retail, commercial, industrial. We have expanded the Snapdragon X Elite into industrial. We've introduced a new category called the on-premise appliance, which we talked about at CES. We are going to be focusing on robotics coming up in the next couple of months, and we have a lot of focus on software and actually SaaS.
I want to talk to you today about a recent acquisition that we just announced yesterday of Edge Impulse. Edge Impulse is a well-known company; it's been in the AI space for a number of years. A few things have been going on, Jim, when it comes to Industrial. As AI has matured so much over the last five years, with so much focus and energy spent on the technology itself, we have many different types of data series—whether it's pressure, temperature, location, humidity, audio, of course, images. You start to extract intelligence out of this data at the edge, you combine it with an LLM, and you can now truly automate a workflow. So the two words that I use when it comes to industrial is situational awareness. Can I get a much better sense of situational awareness to an agent? And when we started to talk to the Edge Impulse guys about five months ago, what we realized was this was a theme that was a combination of embedded AI, ML, and SaaS. The entire toolchain is cloud-based. The idea is for you as a developer to bring your data, find a model, build a model, bring the customer's own model, and then start to figure out what's the best way to extract intelligence from that model. You train, you fine-tune, you keep running it, you get to a point where you have something in the cloud that is running well, then you deploy to the edge. When you deploy to the edge, you're obviously going to have some loss in terms of what the hardware and the edge can support. When we worked on AI Hub last year, a lot of the effort was focused on quantization for what the hardware and the silicon could support. We've had seamless integration. We've had models that we are training in the cloud; when we deploy using AI Hub into the actual device, you're able to see performance, you're able to see what will be the inference per second, what is the level of complexity you can run, what kind of concurrencies you can run.
And then we have a second step. Once you've figured out what the model is, how you're going to run it, there is a full deployment toolchain. How do you actually take that model, put it into a container, connect it to an SDK, and run it in an application? We have announced a VS Code-based IDE, a Qualcomm IDE, that you as a developer can now simply take that model in your application, see it deployed on the IDE, you open the IDE, you start to make code changes, deploy it. It took us about two weeks for us to start to work with the Edge Impulse team for their MLOps IDE to connect to all of our environment. If you zoom in over here for a second, this is essentially taking a camera input or a video input, feeding it into a GStreamer pipeline that is running on our IQ9, which we announced at MWC Austin in October. It goes into our SDK, goes into a model that Edge Impulse has created. You are first doing object tracking. In this case, this is an example of a car wash. Imagine that I am tracking objects, I know this is a car. As soon as I know that, I put a bounding box around it, I feed it into a second model, which is an LLM. The LLM already knows what this car is, so I can run a simple reasoning agent and say, 'Please give me stats on how many cars showed up, what makes and models were there, what colors were there, did they have any damage, what does the front look like, what did the back look like.' This is being created on the cloud, all on the cloud. Once you create that workflow, you simply click deploy. Completed. If I show you, if you walk here this way a little bit, this is a different use case running on the same silicon. I have Dragon Ring, this is our IQ9 EVK. 35 concurrent camera streams at various resolutions. So that same piece of silicon, 35 concurrent streams running on this. Now imagine running that same use case here. This is a smart city application. I want to know frequency of traffic, volume of traffic. Imagine I put a second camera where I need to know with more precision, more accuracy, more resolution, what does this look like. So we are trying to now start to think about this business in terms of what is the silicon that we need to have, T-shirt size the silicon, figure out what is the end use case. Once you understand the use case, think about it in the context of what is the kind of data—is it image data, is it time series data, is it a combination of the two, is there a multimodality aspect of it? Go to Edge Impulse, create that workflow, figure out what are the stages, what are the cascades that you need to be able to get to the output. Once you have the output, deploy it in the application. Let me show you something else at a different scope.
This is an application. There is a surveillance application. This is an application. I have two pairs of cameras there. The ones that are lower down below are legacy cameras; they don't have any intelligence. It's simply an Ethernet stream that goes into an edge box, kind of like what you saw. The cameras are proper smart cameras; these actually have AI running in the cameras. Same exact idea: you're extracting features and you're dealing with them. Let's check this out.
Show me Nike shoes. Wow, that's pretty cool. I like that. So what am I doing here? I'm extracting features from the video stream in real time. I can keep training the model, I can make it better. Once I extract it, I tag that instance with what was relevant information in that window of time. Now I have a time series that is tagged. I can do this across multiple cameras. So I say, 'Hey, backpacks.' Somebody showed up wearing a backpack. Okay, wow. I think the backpack had this toy in the back. Person. Amazing. Can you show me where the location is? So the thing is, is that you would tap into—so the location would be up in this area. Where is that person? That's amazing. So we've gone from all on-device. I want to create tools—this is the SaaS service, yeah, this SaaS service—but I want to create tools that allow me to be able to figure out how to deal with situational awareness. So I have to be very familiar with how to deal with data. It's my data, I want to create an application around it. I need to be able to have a very good understanding of what my hardware is capable of, because the deployment of the hardware is connected to my situational awareness. So something like this, this is a highly complex surveillance service. Oh yeah, I have thousands of cameras that individually have intelligence, either at the edge or that stream is connected into some kind of an edge box, and I'm extracting that specific intelligence. Once I have that, then it's an application that feeds into an LLM. Yeah, right. So for us, this is complete democratization of various types of ML, various types of CV, various types of GenAI coming together to build actual use cases. I want to now show you the on-premise appliance.
This is something that we introduced at CES, right? This is something that will sit on your shop floor. It is something that will sit in a small or medium enterprise. This is running a 100 billion parameter model. This is right now running 670 billion. Wow, yeah. It can also take in an Ethernet stream and run all the CV models that you saw locally here. So I have simple chips, I have slightly complicated chips, I have more complicated chips, and I have a box and a complete hybrid AI system from the edge. In fact, this is being requested by customers who run air-gapped systems. They cannot be connected to the internet; they can't allow anything to come in. So it can run hybrid. You can see use cases here that are truly hybrid. I'm running partly something up here; if it needs to go check back, it goes to the cloud. So what I love about this is this entire journey of AI, hybrid AI, speculative AI—what do you decode? Putting it into practice in the context of what a business actually needs and what it's market. It's very customizable to whatever that business requires, because for us, what we have done is we completely separated the entitlement that the hardware provides to the layer above. If you say, 'Here's my piece of hardware, I have an operating system,' the hardware can run AI on CPU, on GPU, on NPU. They provide different types of entitlement. I have an AI layer that knows how to deal with different types of models. I have a variety of different sensors, camera blocks, etc. Once I can get to the point where I'm extracting intelligence from a piece of time series, now run a service. Service can run locally, can run in the cloud, it can run hybrid, and it's all use case.
Tremendously, tremendously. Look, I think the rate at which AI has—you know, in fact, I would say this: what GenAI did was it allowed intelligence that was very hard to convert into what I consider as human-consumable, it made that change. You know, the idea for a car wash to say, 'I would like to know how many Audis showed up here because somebody forgot their keys in an Audi'—pretty real-world use case, right? Hey, I forgot my wallet and I was driving an Audi. How do I go figure out who to go talk to? Show me the Audis. What color was it? Black. Show me black Audis. Think about the value of that. Could you do that if you only had ML? Would you do that if you only had CV? What GenAI has done is it has made this entire continuum to where you can actually take machine information, put a tag on it, put a strong LLM or LVM on top of it, make it train, you know, write guardrails, rules, etc., and say, 'Okay, now I run a business.' So for us, industrial has been, for me frankly, over the last year or so, the eye-opening journey has been how do you solve people's problems that they have so far been trying to solve by treating the edge as an endpoint to a dumb bit pipe that you stream to the cloud and then it sits in the cloud and you use very expensive resources to go get that information, which frankly doesn't scale. Because AI has changed so rapidly over the last five years, and especially over the last two years, the tools are all available. You know, I mean, what Edge Impulse has created was not created when GenAI was around. It was an MLOps IDE with the mindset that this should be democratized. I feel so great about the ability for us to be able to take a tiny startup, take the technology, take the people, take the intelligence, and then make it available to so many things that we're doing. I've spoken today to companies that build home appliances, dishwashers. I've spoken to companies that build displays for hotels. I've been talking to businesses like a car wash. We talk to security companies. So the takeaway is, it's endless. Like, there is no real end to what is possible.
