Edge AI, where video is analyzed locally closer to the data point, is gaining prominence in security. In this regard, Ambarella has various advanced solutions, including the N1 family of system-on-chips (SoCs) that make edge appliances such as AI boxes more intelligent, energy-efficient and cyber-secure. This article takes a closer look at the N1 family, as well as the new CV7 SoC catering to a variety of camera form factors.
N1 family: Addressing the near-edge tier
The N1 family of SoCs addresses the near-edge tier – the layer where intelligence sits alongside the video management system or the network video recorder. The latest addition to the series, the N1-655, provides on-chip decode of 12 simultaneous 1080p30 streams (1080 resolution at 30 frames per second) and concurrently runs a hybrid of multimodal VLMs and CNNs within approximately 20 watts. The higher-performance N1 SoC operates under approximately 50 watts and supports multimodal vision analysis across up to 32 camera sources.
Applications include edge AI boxes, autonomous mobile robots and smart city security video recorders. Especially, edge AI boxes have gained popularity in security. In video surveillance, many users still use legacy systems where cameras have little to no AI/analytic capabilities. These users want to keep their cameras but are reluctant to stream video to the cloud for analysis. Edge AI boxes, then, present an ideal solution addressing legacy users’ AI processing needs. And with the N1 SoCs, AI boxes become more powerful and intelligent than ever.
“What the box itself can now do is qualitatively different. Forensic search becomes a natural-language conversation across hours of footage, situational awareness extends in real time across multiple cameras concurrently, and the pixels stay inside the customer's network, which improves privacy while reducing bandwidth and cloud cost. The N1 family also gives integrators headroom to add new AI capabilities through software updates, extending the productive life of a deployment and lengthening the return on the initial investment,” said Jerome Gigot, VP of Edge AI Marketing at Ambarella.
As aforementioned, the N1 SoCs have superb compute power and can run multiple AI models, such as CNNs, VLMs and LLMs, at the same time. This is significant as no one model can address all of users’ needs. On the other hand, running different models on one SoC can meet most of the user’s requirements – from detecting abnormalities to setting rules or conducting search using natural language.
“CNNs remain the most accurate option for high-frequency, well-defined tasks like license plate recognition, face matching, and fire-and-smoke detection. VLMs are stronger when the question is open-ended, such as describing a scene, reasoning about anomalies, or responding to natural-language instructions. LLMs add value for summarization, policy reasoning, and operator-facing dialogue. A real-world security deployment needs all three classes available, with the right model routed to the right task,” Gigot said.
Vertical markets that can benefit from the N1 family
According to Gigot, the N1 family fits best where multiple camera streams need to be aggregated and reasoned over in real time, and where the customer has a strong preference for keeping that processing on-premise.
“In smart cities and public safety, that profile fits traffic management centers, transit hubs, and municipal surveillance operations rooms. Enterprise security deployments at large corporate campuses, data centers, casinos, sports venues, and retail flagships fit the same model. Critical infrastructure carries an additional layer of constraint, since airports, ports, energy facilities, and utility substations often operate with limited bandwidth back to a cloud or with compliance rules that restrict what can leave the site, and the N1 family's on-premise model addresses both of those conditions directly,” Gigot said.
CV7: Making cameras smarter and more power-efficient
For cameras that need powerful SoCs to run complex AI models, Ambarella has recently announced CV7, which is an upgrade from the previous CV5 SoC. Built on Samsung's 4-nanometer process – a first for an Ambarella SoC at this node – the CV7 delivers more than 2.5 times the AI performance of the CV5, as well as support for the latest AI architectures including VLMs, LLMs and vision transformers. With power consumption reduced by approximately 20 percent versus the prior generation for the equivalent use case, the CV7 is suitable for a range of applications including enterprise security, drones, sports and wearable cameras and telematics.
In video surveillance applications, the CV7, which sees Ambarella’s third-generation CVflow AI accelerator, image signal processor, hardware video encoder, Arm cores and I/O all integrated onto a single die, can process multiple streams up to 8Kp60 concurrently. Further, it allows a range of advanced camera functionalities including high dynamic range, fisheye de-warping, dual-fisheye merging, and three-dimensional motion-compensated temporal filtering, with low-light sensitivity down to 0.01 lux. These capabilities make the CV7 a strong fit for multi-imager panoramic cameras, multi-sensor PTZs, 360-degree and dual-fisheye designs, high-resolution bullets and turrets used for license plate or face capture at distance, and 8K consumer and prosumer cameras such as action and 360 cameras.
Verticals that can benefit include large transportation hubs, sports and entertainment venues, retail flagship stores, critical infrastructure perimeters, casino floors, and warehouse and logistics environments, where a single multi-sensor camera can now cover ground that previously required several fixed cameras and an external compute appliance.
Similar to the N1 family, the CV7 is designed to run multiple AI models simultaneously including CNNs and transformer networks. “In a mature deployment, the VLM on the camera functions as an open-ended perception and language interface, then hands off to a specialized CNN when a task demands its precision, such as face matching or fire-and-smoke detection. You get general scene reasoning from the VLM and the accuracy you need on the high-frequency tasks from the purpose-trained model. Running both classes inside a single SoC, within the thermal envelope of an IP camera, is exactly what the CV7 is designed to do,” Gigot said.
Key cybersecurity features of N1 family and CV7
Needless to say, edge devices are subject to cyber threats and attacks. Protecting edge devices against these threats therefore becomes critical. In this regard, both the N1 family and CV7 play a key role.
“The CV7 and the N1 family ship with a hardware root of trust supported by Arm TrustZone, secure boot to validate firmware before execution, and secure storage and key provisioning for device credentials. The platform also includes a true random number generator for cryptographic operations, one-time programmable memory for device identity and feature provisioning, and DRAM scrambling and virtualization to protect against memory-bus probing and inter-process leakage,” Gigot said. “Asymmetric and symmetric crypto acceleration is handled in hardware, which keeps TLS, secure firmware updates, and encrypted communications off the application CPU.”
Gigot adds that integrating the ISP, AI accelerator, video encoder, and security functions on a single die also reduces the attack surface that multi-chip designs introduce through external buses. “In practice, this stack gives camera and appliance makers what they need to meet evolving regulatory and procurement requirements, including secure-by-design specifications from enterprise and government buyers, and it supports the kind of over-the-air update model that field-deployed devices require to stay current over their operational life,” Gigot said.