Edge AI demands devices that can think locally and act instantly. In an exclusive interaction, Muneyb Minhazuddin and Jerome Gigot from Ambarella tell EFY’s Akanksha Sondhi Gaur how their CV7 SoC, multi-sensor fusion, and Developer Zone strategy empower cameras, robotics, and automotive systems, enabling startups and OEMs to deploy high-performance AI solutions at the edge.
Q. Can you give us an overview of Ambarella’s presence and focus in the edge AI and IoT space?
A. We have been a leading player in video processing and edge AI solutions for over 21 years. We began with broadcast and video networks, and today we focus on IoT endpoints, automotive, robotics, and edge appliances. Our portfolio spans hardware, software, and services, enabling devices—from cameras to robots—to perform perception, inference, and AI tasks locally. Our strategy is to offer a consistent full stack across all system-on-chip (SoC) families, making application development portable and efficient.
Q. Tell us about your core strategic focus and market positioning. How has it evolved?
A. We began by designing highly efficient SoC video processors for cameras. Over the past few years, this foundation has expanded into a broader edge AI vision that combines silicon and software to support advanced perception, multi-sensor fusion, and embedded inference at the edge.
We now address three key segments: power-constrained endpoints such as cameras and IoT devices; near- and far-edge appliances that aggregate data and run more complex AI workloads; and physical AI systems, including robots, drones, and vehicles, that require real-time perception and action. Across all segments, our strategy is centred on delivering a complete, tightly integrated stack that combines silicon, software, and developer enablement, purpose-built for edge deployments rather than cloud or datacentre environments.
Q. What are the key markets you are pursuing today?
A. Our primary focus spans enterprise and smart security cameras, sports and wearable cameras, robotics and autonomous machines, automotive perception and advanced driver-assistance systems (ADAS), and edge AI appliances and gateways. We differentiate through power-efficient silicon with robust real-time performance, supported by a strong developer ecosystem that accelerates time-to-market.
Q. Where do you see the biggest growth opportunities in the next 3–5 years?
A. We expect rapid growth in multi-sensor fusion platforms that combine cameras, radar, lidar, and IMUs; transformer-based and multi-modal AI models running at the edge; and embedded GenAI applications where local inference and data privacy are critical. Automotive ADAS platforms with robust vision and perception stacks are also gaining momentum, driven by the need for local intelligence, reduced reliance on the cloud, and real-time responsiveness.
Q. How do you balance silicon innovation with software and ecosystem investments?
A. Silicon is foundational, but without proper tooling, models, and developer workflows, solutions cannot be deployed efficiently. We therefore invest in a consistent SDK and toolchain, pre-optimised AI models and reference designs, and strong developer productivity through the Developer Zone, alongside partnerships with AI framework vendors and edge AI ISVs. This approach ensures that customers can move from concept to product quickly and reliably.
Q. What do you think about partnerships versus building in-house capabilities?
A. We develop core video, perception, and real-time inference capabilities in-house, while partnerships with framework providers and ISVs expand model ecosystems and agentic workflows. For strategic technologies, such as proprietary radar processing, we have pursued acquisitions, including Oculii, to integrate critical capabilities directly into our stack.
Q. What is your global go-to-market strategy?
A. We engage directly with OEMs and Tier-1 partners, extend our reach through distributors and ecosystem partners, and enable independent developers and startups via the Developer Zone. Systems integrators contribute vertical application expertise, complementing our silicon and software capabilities.
Q. How important is India as a market for you?
A. India is strategically important due to growing adoption of edge AI in smart cities and public-safety systems, a rapidly expanding automotive and EV market, and strong government support for electronics manufacturing and innovation. India’s engineering talent pool and startup ecosystem align closely with our edge-focused silicon and software strategy.
Q. What traction have you seen in India so far?
A. We are seeing early traction in enterprise security and CCTV markets. We are among the first to achieve STQC certification for cameras built on our platforms, and we have announced design engagements with Honeywell India for smart surveillance products. These developments reflect strong trust from Indian OEMs handling demanding vision workloads.
Q. Beyond the listed wins, what additional India partners and certifications exist, and what is the expansion timeline?
A. We partner with VVDN, Sparsh (first STQC-certified CCTV range), Honeywell, eInfochips (an Arrow-acquired design house with an India presence), and eCon Systems (robotics modules showcased at CES). While we do not yet have a direct office, we aim to establish a GTM and R&D presence and deliver ‘Made-for-India’ products by the end of 2026 through strategic engagements.
Q. How do you plan to expand in the Indian automotive market?
A. Automotive traction in India has been slower than in China, Europe, and North America, but momentum is building. We are actively engaging with local Tier-1 suppliers and OEMs on ADAS and surround-view systems, aligned with India’s functional safety standards and vehicle electrification trends. Our approach includes collaborating with local suppliers, developing reference designs tailored to regional use cases, and supporting relevant safety and compliance requirements.
Q. Will Ambarella India have localised R&D or co-innovation programmes?
A. Yes, we see strong value in engaging Indian developers and startups through local developer communities, hackathons, and academic collaborations. Over time, localised R&D initiatives will tailor edge AI solutions to Indian use cases, aligned with market maturity and adoption.
Q. What industry partnerships in India are crucial to your strategy?
A. We are strengthening ties with OEMs, system integrators, distributors, government certification bodies, and independent software vendors focused on edge AI. These partners help translate our silicon capabilities into deployable products across infrastructure, enterprise, and automotive segments.
Q. You recently announced the CV7 SoC. What are its core highlights?
A. CV7 is our third-generation edge AI SoC, built on a 4 nm process node, enabling ultra-low-power 8K video processing with multi-sensor fusion and on-chip AI. It supports a broad range of applications, including sports and wearable cameras with AI-based on-device editing and low-light enhancement; security cameras with on-device video understanding and vision-language model (VLM) support; automotive ADAS with localised processing for features such as automatic braking and driver monitoring; and multi-camera robotics applications that stitch multiple 4K sensors or fuse LiDAR and radar locally.
Q. What architectural innovations enable it to handle ultra-low-power 8K video and AI workloads?
A. Our design is algorithm-first. We optimise memory access, internal pipelines, and bus architecture to minimise off-chip memory usage. This allows multi-camera stitching, encoding, and AI inference to run simultaneously without relying on generic GPUs that repeatedly access external memory. On-chip AI models load efficiently, reducing latency and power consumption.
Q. How does it manage memory bandwidth, caching, and power to prevent thermal throttling?
A. It uses high-frequency LPDDR5 to ensure efficient memory throughput. Moving from 5 nm to 4 nm reduces power consumption by approximately 20% for the same workload. Unlike smartphones that throttle dynamically, we guarantee consistent performance for 24/7 applications such as cameras and automotive systems by carefully balancing memory, compute, and thermal design.
Q. How does it differ from GPU-based or purely NPU-centric designs?
A. Unlike generic GPUs, which process workloads sequentially and repeatedly access memory, we tightly integrate the ISP, AI accelerator, and CPU clusters for real-time parallel processing. The architecture is tailored for vision workloads, multi-camera pipelines, and transformer-based AI models such as VLMs and lightweight LLMs, delivering up to 2.5× higher AI performance than the previous generation while maintaining low power consumption.
Q. Which AI models run natively on CV7, and how are they optimised?
A. CV7 supports ResNet, MobileNet, Vision Transformers, CLIP, LLaVA, and other transformer-based models. Optimisations include pruning, quantisation, and memory-efficient execution, leveraging on-chip AI acceleration to handle multi-modal workloads across image, text, audio, and radar inputs.
Q. How is multi-modal data fused for detection, segmentation, and tracking?
A. We support both low- and high-level sensor fusion. Input streams—video, audio, LiDAR, and radar—are tokenised and processed through our transformer-based AI engine. The hardware efficiently ingests these inputs, applies AI locally, and generates real-time outputs for robotics, automotive, or surveillance applications.
Q. How does the Image Signal Processor integrate with AI inference pipelines?
A. The ISP outputs feed directly into AI accelerators at multiple processing stages, even before final colour correction. This tightly coupled flow enables AI-assisted image enhancement, motion estimation, and predictive denoising.
Q. From a compiler and scheduling perspective, how are mixed workloads handled?
A. We use a graph-based compiler to analyse models at compile time. It partitions layers, schedules memory access, and optimises buffer reuse to avoid runtime stalls or unpredictable delays. Memory allocation is largely static for large transformer models to guarantee worst-case performance, favouring predictability over flexibility.
Q. What firmware and software priorities ensure deterministic, low-latency performance?
A. We provide Linux- and RTOS-based firmware depending on the application. Security cameras use Linux, while drones, robots, and automotive ADAS platforms use RTOS options such as QNX or ThreadX. Firmware schedules tasks, manages sensor data efficiently, and ensures real-time AI inference without dropped frames.
Q. How does the Oculii acquisition enhance CV7’s multi-modal fusion for ADAS?
A. Oculii’s AI radar software enables low-level sensor fusion of radar, LiDAR, camera, and inertial measurement unit inputs, training networks for precise object detection rather than inefficient high-level matching. Combined with CVflow 3.0, transformers and multi-modal workloads execute natively for vision-language and large language models, enabling real-time navigation for drones and vehicles.
Q. What improvements does CV7 bring over the previous generation?
A. It delivers a 2–3× improvement in AI performance, multi-camera processing, and power efficiency. It supports 8K video at 60 fps, multi-sensor fusion, transformer-based AI, and 4 nm fabrication for longer battery life and a lower thermal footprint. These capabilities extend its use across enterprise security, sports cameras, drones, robotics, and automotive ADAS.
Q. What is the timeline for CV7-based products?
A. Customers will begin design-in shortly after CES, with commercial products expected from late 2026 through 2027, including high-resolution security cameras, sports cameras, and automotive ADAS systems.
Q. Can you explain the developer ecosystem Ambarella is building?
A. We provide early access to ISVs to leverage our SoC and software stack. We offer pre-trained AI models fine-tuned for our chips, no-code and low-code orchestration tools, blueprints for edge applications in smart cities, transportation, and industrial automation, and support for frameworks such as PyTorch and ONNX. This enables startups and OEMs to build AI applications without deep expertise in our software development kits.
Q. How will this help Indian startups and OEMs?
A. Our Developer Zone lowers barriers for companies without specialised AI teams. Startups can quickly ingest data, deploy models, and optimise workflows using pre-built tools and examples, enabling faster time-to-market and supporting the growth of an edge AI ecosystem in India.
Q. How do debugging, profiling, and real-time AI visualisation tools help developers?
A. Early access includes profiling, optimisation, and deployment tools. Developers can containerise applications, deploy efficiently, and monitor performance. Observability and lifecycle management support rapid iteration and reduce friction during deployment.
Q. Where can developers access models and tools for early engagement?
A. Early access is available through the Developer Zone, where selected ISVs can test models, build applications using low-code and no-code workflows, and optimise SoC performance. The portal provides pre-tuned models, example use cases, and orchestration tools.
Q. How do you ensure long-term ecosystem support for developers?
A. By exposing our software stack, SDKs, and tools through the Developer Zone, we enable developers to iterate quickly, deploy reliably, and maintain performance. With over 400 million SoCs shipped and a consistent software platform, developers benefit from faster time-to-market and dependable deployment support.
Q. How does Agentic AI differ from physical AI and traditional perception?
A. Physical AI uses multiple sensors for perception and action. Perception AI combines sensor data to interpret environments for robotics or navigation. Agentic AI adds automation and orchestration, where each module or sensor functions as an independent agent that dynamically builds workflows using no-code capabilities. This supports on-device reasoning and action across sectors such as warehousing, industrial automation, consumer robotics, and defence.
Q. How do you see AI evolving from cloud-dependent systems to fully local intelligent devices?
A. We foresee a hybrid model in which real-time decision-making happens at the edge, while metadata learning and broader pattern recognition occur through the cloud. Edge AI will increasingly rely on smaller, domain-specific models optimised for cameras, vehicles, and robots, enabling faster and more reliable real-time action.
Q. How do you approach functional safety?
A. CV7 supports secure boot, memory protection, TrustZone, and hardware isolation. For automotive workloads, these capabilities enable ISO 26262 compliance when integrated with host processors.
Q. How does security factor into the silicon and software architecture?
A. We integrate secure boot chains, encrypted model storage, memory scrambling, and root-of-trust mechanisms to protect intellectual property and prevent unauthorised access to models and firmware.
Q. What is next beyond CV7?
A. Future devices will further improve AI scalability, energy efficiency, and multi-sensor fusion while deepening ecosystem integration. We are also expanding software services, particularly in continuous deployment and remote manageability.
Q. Final advice for design engineers considering Ambarella for their next product?
A. Understand the system-level use case, including sensor choices, AI requirements, and power constraints, and engage with our tools and ecosystem early. Early profiling and integration lead to more predictable performance and smoother field deployment.