Ceva's NeuPro NPUs Double AI Throughput and Boost Efficiency for Multimodal Workloads

Ceva, Inc. has unveiled its next-generation NeuPro™ NPU portfolio, designed to enhance Physical AI applications. The new portfolio boasts up to twice the AI throughput per mm² and a 50% improvement in energy efficiency compared to its predecessor. This advancement is particularly geared towards supporting demanding multimodal, generative, and agentic AI workloads, extending from always-on sensors to advanced edge systems and even into AI infrastructure. This development is significant for practitioners in cloud, DevOps, and AI because it directly addresses the computational and power constraints often encountered when deploying advanced AI models, especially multimodal ones, at the edge. As AI applications become more sophisticated, requiring the processing of diverse data types like text, images, audio, and video simultaneously, the need for efficient, high-performance hardware outside of large data centers becomes paramount. This release enables developers to build more capable and responsive AI systems that can operate closer to the data source, reducing latency and bandwidth requirements. The trend towards hybrid AI, where intelligence is distributed across devices, edge systems, and the cloud, is well-established. Multimodal AI itself has moved from an emerging capability to an expected baseline for serious AI systems in 2026, with models that process only text increasingly viewed as incomplete. The challenge has shifted from merely making AI multimodal to effectively deploying these complex models in real-world scenarios with practical constraints. Ceva's new NPUs fit squarely into this trend by providing the underlying hardware necessary to power this distributed intelligence, particularly for tasks requiring real-time perception and response. In practice, this means that developers working on applications like autonomous systems, robotics, and advanced edge devices will find it easier to integrate sophisticated multimodal AI capabilities. The improved efficiency translates to longer battery life for portable devices and lower operational costs for edge deployments. Practitioners should investigate how these new NPUs can be leveraged within their existing or planned hardware architectures, especially for workloads involving real-time processing of multiple data streams. Understanding the trade-offs between on-device processing and cloud-based inference will be crucial, and Ceva's offering provides a stronger foundation for pushing more intelligence to the edge. This also signals a broader industry move towards specialized hardware that can handle the unique demands of multimodal and agentic AI, moving beyond general-purpose compute.