Ceva has secured the first production deployment of its NeuPro neural processing unit family in Mobilint’s REGULUS edge AI system-on-chip. The South Korean semiconductor company combines Ceva’s NeuPro-Nano NPN32 NPU and SensPro AI digital signal processor with its own chip architecture to handle sensing and neural inference in embedded devices.
Mobilint specifies REGULUS at 10 trillion operations per second (10 TOPS) of AI performance at under 3W, with support for up to 8GB of memory. The processor is aimed at compact systems that process incoming data locally, including robotic equipment, industrial cameras, drones and connected devices operating within constrained electrical and thermal budgets.
Raw camera images and other sensor measurements generally need preparation before an AI model can interpret them. SensPro provides programmable digital signal processing for those inputs, while NeuPro-Nano accelerates neural-network inference. Allocating work between the engines gives equipment designers scope to balance signal processing, recognition and memory movement within one silicon platform.
The distinction between engines also affects software design. Filtering, image correction and signal transformations can run on a programmable processor, while repeated neural-network calculations can use a dedicated accelerator. Practical latency and energy use depend on how the application moves data between the engines, which numerical precision it uses and whether the selected model fits the available memory.
The headline 10 TOPS figure describes a stated peak computing capability rather than a measured number of products inspected or robot actions completed. Camera frame rates, preprocessing, model architecture, transfer overhead and the timing of outputs all influence throughput in a deployed system. Equipment developers will need application tests to establish how the complete pipeline performs.
Power consumption has particular importance in distributed edge installations. A processor attached to a camera above a production line, carried by a mobile robot or fitted inside a compact sensor enclosure may have limited cooling and an established power budget. Reducing the need for remote computation can also simplify operation where connectivity is intermittent, although many devices will still exchange data with supervisory systems for configuration, reporting and maintenance.
Ceva supplies licensed processing intellectual property embedded inside a Mobilint-designed commercial chip. REGULUS is therefore Mobilint’s system-on-chip, with Ceva reporting its first production use of the NeuPro NPU family through that product. The resulting capabilities depend on the configuration of the finished board, sensors and software as well as the chip itself.
Mobilint already ships the MLD-R1 USB AI accelerator built around REGULUS. First shown at COMPUTEX 2026, it enables customers to add local AI processing to an existing host system without immediately designing a new circuit board. Mobilint lists computer vision, video analytics and generative AI among the supported workload types, giving developers an available evaluation route for the architecture.
Software support spans PyTorch, ONNX, TensorFlow, TensorFlow Lite, Keras and Hugging Face ecosystems. Compatibility with those development environments can help move trained models towards the device, although deployment still requires attention to operator support, numerical format, memory demand and integration with host applications.
Industrial image analysis commonly combines deterministic measurement and learned recognition. A camera system might check whether a label is present, identify a surface defect and communicate a result to an inspection controller. Those tasks can differ considerably in timing requirements. A missed frame may be acceptable during an occasional manual review but unsuitable on a fast moving line where the result must reach the reject mechanism before the product passes it.
Separate signal-processing and neural engines can distribute work along this inspection sequence, but camera capture, image conversion, inference, decision reporting and actuator timing all contribute to end-to-end latency. Production engineers must validate that complete path against the position and speed of products travelling through a line.
Robotic applications add further demands because a machine has to act on new information while its surroundings change. Continuous inference can support object recognition or scene interpretation, but the robot’s motion controller, localisation and functional safety systems still need independent engineering. Ceva and Mobilint are presenting processing capacity for these applications rather than a finished autonomy or safety-certified system.
Running inference near industrial equipment can reduce the need to transmit raw sensor streams, particularly where a device must respond locally or external connectivity is limited. The balance between local computing and central supervision will depend on the model size, cybersecurity controls, maintenance arrangements and records required by each application.
REGULUS is already in commercial silicon, and the MLD-R1 accelerator is shipping. Further evidence of industrial adoption will come from equipment suppliers integrating the chip into production machinery and demonstrating reliable application performance under their own power, timing and environmental requirements.
