CEVA Unveils NeuPro‑S: Next‑Gen AI Processor for Edge Deep Neural Network Inference

CEVA, a global leader in digital signal processing, today introduced NeuPro‑S, its next‑generation AI processor architecture designed for deep neural network inference at the edge.

Paired with the industry‑first CDNN‑Invite API, a deep‑learning compiler that orchestrates heterogeneous co‑processing of NeuPro‑S cores alongside custom neural‑network engines, the platform delivers a unified, runtime‑optimized firmware for vision‑centric devices.

NeuPro‑S is engineered for tasks such as segmentation, detection and classification in video and image streams, offering system‑aware enhancements that reduce external SDRAM traffic, compress model weights, and scale across multiple CEVA‑XM6 DSPs, NeuPro‑S cores and bespoke AI engines within a single architecture.

Compared to CEVA’s first‑generation processor, NeuPro‑S delivers on average 50 % higher performance, 40 % lower memory bandwidth consumption and 30 % lower power draw.

The family comprises three pre‑configured models:

• NPS1000 – 1,000 8‑bit MACs per cycle
• NPS2000 – 2,000 8‑bit MACs per cycle
• NPS4000 – 4,000 8‑bit MACs per cycle, achieving up to 12.5 TOPS at 1.5 GHz and scalable to 100 TOPS.

Designed with automotive safety in mind, NeuPro‑S complies with IATF 16949, ISO 26262, and A‑Spice, making it ready for safety‑critical deployments in autonomous vehicles, smartphones, surveillance cameras, consumer cameras, AR/VR headsets, robotics and industrial automation.