IN Brief:
- Mobilint’s REGULUS SoC marks the first production deployment of Ceva’s NeuPro neural processing IP.
- The device integrates a NeuPro-Nano NPN32 NPU and SensPro AI DSP and is specified at 10 TOPS below 3W.
- Mobilint’s MLD-R1 USB accelerator is already shipping, providing a commercial evaluation route for the architecture.
Ceva has secured the first production use of its NeuPro neural-processing intellectual property in Mobilint‘s REGULUS edge AI system-on-chip. The Korean chip developer integrates Ceva’s NeuPro-Nano NPN32 neural processing unit and SensPro AI digital signal processor into REGULUS, alongside its own processing technology. The resulting device is intended to handle sensor data and inference locally in embedded equipment.
Mobilint specifies REGULUS at 10 trillion AI operations per second, with a power figure below 3W and support for up to 8GB of memory. It is designed to operate as a standalone AI processor rather than relying on a larger host for every inference task. The published numbers describe the company’s platform specification; achievable throughput and energy use in a finished product will vary with model size, numerical precision, memory traffic, and the surrounding electronics.
The combined architecture places different stages of computation on specialised engines. SensPro is a programmable DSP for sensing and signal processing, including operations that prepare data for later analysis. NeuPro-Nano provides dedicated acceleration for neural inference. Mobilint’s own AI processing resources complete the configuration, allowing developers to arrange their workloads according to the capabilities of each processing block.
In a machine-vision system, camera data may need to be acquired, transformed, filtered, and reformatted before inference begins. Running suitable stages on the DSP can free neural hardware for models that classify objects or extract other features, while the rest of the device handles control and communications. The exact benefit depends on software partitioning and the interfaces available to the sensors, so the architecture’s flexibility must be exercised in the application design.
Memory bandwidth and data movement constrain edge inference as much as nominal arithmetic performance. Neural weights, intermediate outputs, and incoming sensor samples compete for storage and access to the compute units. The 8GB supported-memory ceiling can accommodate a range of workloads, but a product designer still has to configure the memory system, power supply, peripheral connections, and software scheduling so that data reaches the engines when required.
Mobilint has taken REGULUS into a shipping product through its MLD-R1 USB AI accelerator, previously shown at COMPUTEX 2026. This gives developers a commercial route to evaluate the SoC alongside existing computers and embedded hosts. Its availability gives embedded developers a production component built around Ceva’s licensed neural-processing technology.
Mobilint identifies computer vision, video analytics, and generative AI among its target workloads, with potential deployments in industrial inspection, robotics, security equipment, smart cameras, and other edge devices. Supported development workflows include frameworks and model formats familiar to AI developers, although a trained model still has to be compiled or mapped to the available operators, precision modes, and memory configuration before it can run efficiently on an accelerator.
The mapping process also affects work carried out by the DSP. Signal-processing algorithms and neural models may have different latency requirements, and a camera or other sensor can continue delivering samples while an earlier frame is being analysed. Coordinating those stages is necessary to avoid dropped data or unpredictable response times. The presence of a programmable signal processor adds a resource for operations that do not map naturally onto a neural network engine.
Ceva licenses processor and signal-processing designs that chip manufacturers embed in their own devices. Production use of NeuPro-Nano therefore establishes a commercial implementation of the IP family, while REGULUS remains Mobilint’s SoC with its own architecture and product support. The two companies’ technologies contribute different parts of the sensing and inference pipeline.
Further integrations will be determined by Mobilint’s device roadmap and customer requirements for camera inputs, model performance, power consumption, and software support. For now, the shipping MLD-R1 offers a practical platform for developers assessing whether REGULUS can perform the sensing and inference work needed by their particular embedded systems.
