REGULUS brings AI closer to cameras and robots
· BehindLines newsroom
On November 2, 2026, Ceva said that Mobilint had incorporated Ceva’s SensPro and NeuPro-Nano computing designs into REGULUS, a chip intended for embedded artificial intelligence. The supplier described this as the first time its NeuPro family of neural processors had reached a production product. The development provides a commercial reference for technology offered to chip designers. The performance figures remain claims made by the manufacturers.
According to Ceva, REGULUS delivers 10 TOPS of computing performance while consuming less than 3 watts, and supports as much as 8 GB of memory. Its architecture divides work among several engines: SensPro processes signals and sensor data, while NeuPro-Nano handles neural inference, meaning the execution of models that have already been trained. The intended applications include industrial vision, cameras, drones and robots.
On its product page, Mobilint specifies that the 10 TOPS figure refers to dense computation using 8-bit integers, known as INT8. REGULUS can run Linux and an application without a host computer. The manufacturer also describes hardware video encoding at 4K and 60 frames per second, together with direct connections for cameras, sensors, Ethernet and USB. These features place the chip within a complete equipment design.
The move into a product is illustrated by the MLD-R1, a USB accelerator that uses REGULUS. Previewed at COMPUTEX 2026, it is now being shipped, Ceva says. Mobilint also offers a module and an evaluation kit. Its software environment compiles PyTorch, TensorFlow or ONNX models for its hardware; the company says more than 490 architectures have been validated. For an integrator, these provide starting points for trying out applications of their own.
Both sources come from commercial partners. Several limitations nevertheless frame these announcements. The 10 TOPS figure does not measure the speed of a particular application; the consumption of less than 3 watts does not describe the power used by an entire robot. Validation of architectures does not guarantee that every model will be compatible. Finally, the shipping announcement specifies neither volumes nor customers. Tests using the kit will therefore need to establish performance, power consumption and compatibility for each application.