Smart edge technology firm Ceva has introduced Ceva-NeuPro-Nano, a family of very low power Edge AI neural processing units (NPUs). These NPUs allow semiconductor companies and OEMs to integrate TinyML AI models into their SoCs for consumer, industrial, and general-purpose AIoT products.

TinyML enables the deployment of machine learning models on low-power, resource-constrained devices. It allows AI to be incorporated into IoT and embedded devices.

According to research firm ABI Research, by 2030 over 40% of TinyML shipments will be powered by dedicated TinyML hardware rather than all-purpose MCUs. By addressing the specific performance challenges of TinyML, the Ceva-NeuPro-Nano NPUs aim to make AI ubiquitous, economical and practical for a range of use cases, spanning voice, vision, predictive maintenance, and health sensing in consumer and industrial IoT applications.

The Ceva-NeuPro-Nano Embedded AI NPU architecture is fully programmable. It is designed to execute Neural Networks, feature extraction, control code and DSP code, and supports most machine learning data types and operators including native transformer computation, sparsity acceleration and fast quantisation.

According to Ceva, Ceva-NeuPro-Nano NPUs delivers superior power efficiency, with a smaller silicon footprint, and optimal performance compared to the existing processor solutions used for TinyML workloads which utilise a combination of CPU or DSP with AI accelerator-based architectures. Ceva-NetSqueeze AI compression technology directly processes compressed model weights, without the need for an intermediate decompression stage. This enables the Ceva-NeuPro-Nano NPUs to achieve up to 80% memory footprint reduction, solving a key bottleneck inhibiting the broad adoption of AIoT processors today.

“Ceva-NeuPro-Nano opens exciting opportunities for companies to integrate TinyML applications into low-power IoT SoCs and MCUs and builds on our strategy to empower smart edge devices with advanced connectivity, sensing and inference capabilities. The Ceva-NeuPro-Nano family of NPUs enables more companies to bring AI to the very edge, resulting in intelligent IoT devices with advanced feature sets that capture more value for our customers,” said Chad Lucien, vice president and general manager of the Sensors and Audio Business Unit at Ceva. “By leveraging our industry-leading position in wireless IoT connectivity and strong expertise in audio and vision sensing, we are uniquely positioned to help our customers unlock the potential of TinyML to enable innovative solutions that enhance user experiences, improve efficiencies, and contribute to a smarter, more connected world.”

According to Paul Schell, Industry Analyst at ABI Research, “Ceva-NeuPro-Nano is a compelling solution for on-device AI in smart edge IoT devices. It addresses the power, performance, and cost requirements to enable always-on use-cases on battery-operated devices integrating voice, vision, and sensing use cases across a wide array of end markets. From TWS earbuds, headsets, wearables, and smart speakers to industrial sensors, smart appliances, home automation devices, cameras, and more, Ceva-NeuPro-Nano enables TinyML in energy constrained AIoT devices.”

The Ceva-NeuPro-Nano NPU is available in two configurations – the Ceva-NPN32 with 32 int8 MACs, and the Ceva-NPN64 with 64 int8 MACs, both utilising Ceva-NetSqueeze for direct processing of compressed model weights. The Ceva-NPN32 is optimised for most TinyML workloads targeting voice, audio, object detection, and anomaly detection use cases. The Ceva-NPN64 provides 2x performance acceleration using weight sparsity, greater memory bandwidth, more MACs, and support for 4-bit weights to deliver enhanced performance for more complex on-device AI use cases such as object classification, face detection, speech recognition, health monitoring, and others.

The NPUs are delivered with an AI SDK – Ceva-NeuPro Studio – which is a unified AI stack that delivers a common set of tools across the entire Ceva-NeuPro NPU family, supporting open AI frameworks including TensorFlow Lite for Microcontrollers (TFLM) and microTVM (µTVM).