How Agentic AI Transforms Edge Embedded Development

Agentic AI is emerging as a powerful tool to tackle one of the biggest challenges in tech: getting software to run efficiently on specialized hardware at the network’s edge. According to Muneyb Minhazuddin, Customer Growth Officer at semiconductor company Ambarella, this new approach is dramatically simplifying a traditionally complex process.

The Traditional Hurdle: Months of Complex Development

For years, chipmakers like Ambarella relied on a standard playbook to help software companies use their hardware. This involved providing application programming interfaces (APIs), publishing code on platforms like GitHub, running training sessions, and fostering developer communities. “That’s a lot of resources, a lot of people, a lot of money, and a lot of time,” Minhazuddin explained.

The goal was to enable Independent Software Vendors (ISVs) to “port” or adapt their applications to run on specific chips. This process could often take several months, creating a significant barrier to innovation and deployment.

The Agentic AI Solution: From Months to Days

Ambarella’s breakthrough was to wrap an agentic AI layer around its core software tools. Think of this layer as an intelligent assistant that understands both the developer’s goals and the intricacies of the hardware. Instead of developers manually navigating complex SDKs (Software Development Kits), they can now interact with this AI agent.

The results have been striking. ISVs are now able to port their code to Ambarella’s hardware in a matter of days. “[At first, we thought] it must be down to the ISV’s smart engineers, or the technology, or our products,” Minhazuddin said. “But using the same formula with subsequent ISVs, it’s the same thing… This actually works.”

Why Embedded Software is the Perfect Fit for Agentic AI

Unlike the vast, open-ended problems tackled in massive cloud data centers, embedded software for edge devices often has a more focused job. “The good news is, at the edge, it’s very fixed-function,” Minhazuddin noted. This means the tasks are well-defined, like analyzing a video feed from a security camera or processing sensor data in a robot.

Because these functions are narrower, they can be more easily turned into “skills” for an AI agent to manage. This makes the embedded software sector a prime candidate to be one of the first to fully embrace agentic AI for development.

Real-World Impact: Speeding Up AI in Retail and Beyond

Ambarella has successfully onboarded ISVs in healthcare, retail, and robotics, some in as little as three days. This speed is now translating into real-world deployments. \”In the last four months, [our ISV partners] have put our development kits into 20 large retail chains in the United States, large chains of coffee shops, and drive-throughs,\” Minhazuddin shared.

A key trend in retail is hardware consolidation. Stores typically have separate boxes for networking, point-of-sale systems, and now AI analytics. Ambarella’s technology asks: can these three boxes become one? Consolidating these workloads onto a single, powerful Ambarella SoC (System on a Chip) could lead to major cost savings, especially for chains with thousands of locations.

The Hardware Shift: Rethinking CPUs for AI Agents

Balancing Compute Power

Running AI agents directly on edge devices requires a shift in hardware design. In traditional AI processing, a GPU or special accelerator does the heavy lifting, guided by a CPU. With agentic AI, the CPU itself becomes much more important, as it needs to run the agent’s “brain” or orchestration engine.

“The CPU-GPU ratio is changing fast because of agentic AI,” Minhazuddin observed. Ambarella’s current chips, which already contain an Arm CPU core alongside AI accelerators, are naturally suited for this balance. The company is considering future products that further optimize this ratio for agentic applications.

Distributing Intelligence: Cameras and Edge Boxes

Intelligence can be distributed between the camera sensor itself and a nearby edge computing box. Ambarella’s N1-655 chip, used in such edge boxes, is powerful enough to handle local processing, inference, and agentic tasks. “Camera-box combinations can eliminate requirements for big servers,” Minhazuddin said.

The decision of how to split the workload is up to the customer. Ambarella’s strategy is to continue adding more processing power (\”oomph\”) directly into cameras, enabling smarter aggregation of data streams without needing a large server rack.

The Evolving Landscape: Edge to Cloud and Physical AI

Minhazuddin sees the edge AI market evolving in two main directions: from the cloud to the edge, and from the edge to the cloud.

• Cloud to Edge: Large data center operators are moving AI inference tasks closer to where data is generated, using powerful on-premise hardware.
• Edge to Cloud: This includes two exciting areas:
  1. Physical AI: Systems that use data from multiple sensors (cameras, lidar, etc.) to understand and interact with the physical world autonomously, like robots or advanced driver-assistance systems. These systems must make decisions instantly, without waiting for a cloud connection.
  2. The Edge Box: A consolidation point that aggregates and processes data from many sensors, running multiple AI models without needing to send everything to the cloud.

Ambarella’s N1-655 is targeting these edge-to-cloud segments. While multi-chip systems are a future possibility, the focus remains on ultra-efficient single-chip solutions for now.

The Road to Equilibrium

The journey of AI workloads has moved from the data center, toward the cloud, and is now firmly arriving at the edge. Minhazuddin predicts a future equilibrium. “Some workloads run best in the cloud, some in the data center, and some work best at the edge, but they’re all talking to each other,” he said. “That’s where we’ll hit the equilibrium, but I don’t think we’re there yet.”

The adoption of agentic AI for embedded software development is a significant step toward that future, making powerful edge intelligence faster and easier to deploy than ever before.

Article by Sally Ward-Foxton, covering AI and semiconductor technology.