Ceva Expands NeuPro NPU Portfolio into AI Infrastructure

Ceva Inc. has unveiled its next-generation NeuPro neural processing unit (NPU) portfolio, targeting AI workloads ranging from always-on sensing and advanced edge systems to cloud and data center infrastructure. Built on a common architecture and unified software development kit (SDK), the new portfolio delivers up to 2× higher AI throughput per square millimeter and 50% better energy efficiency than the first generation.

The architecture is designed for increasingly demanding multimodal, generative, and agentic AI workloads, including physical AI applications that combine sensor-based perception, real-time responses, and more complex reasoning across devices, edge systems, and cloud platforms. A unified development environment lets semiconductor designers select the right NPU for each workload or integrate multiple processing engines into a single system-on-chip (SoC) while reusing software and development tools.

The new architecture supports mixed-precision AI computation, including FP4 and FP8 formats, to improve inference efficiency and reduce memory requirements and data movement. Higher memory bandwidth, dynamic tensor shapes, and paged key-value (KV) caching address the demands of transformer models and longer context windows. Additional capabilities include optimized tensor reshaping to improve multiply-accumulate utilization, support for Mixture-of-Experts and linear-attention networks, and deployment of pre-quantized models without a separate quantization and calibration step. Compression, quantization, and pruning further reduce computational workloads and data transfers.

The portfolio comprises four NPUs. NeuPro Nano targets always-on sensing and entry-level perception in resource-constrained devices, while NeuPro Prime addresses mainstream convolutional neural network (CNN) and transformer workloads. NeuPro Max targets advanced edge AI, including robotics, multimodal models, on-device assistants, and agentic applications. NeuPro Ultra extends the architecture into AI infrastructure, supporting large language model inference and intelligent processing within memory, connectivity, and data-processing subsystems.

Ceva's expansion reflects the growing importance of memory bandwidth and data movement in large-scale AI inference. NeuPro Ultra combines multicore processing, multi-user operation, high memory bandwidth, and advanced memory management to support cloud and data center deployments. It can also bring neural processing closer to memory and networking components, reducing unnecessary data transfers and easing system bottlenecks.

According to ABI Research, cloud AI inference capacity is forecast to grow at a compound annual growth rate of 42% through 2035. Ceva says leading semiconductor and system companies are already evaluating and integrating the new NeuPro technology. Select early-access customers can access the architecture, with general availability planned for the second quarter of 2027.

Ceva will present its portfolio and physical AI strategy at technology symposiums in Hsinchu, Taiwan, on November 2, and Shanghai, China, on November 5, 2026.

The post Ceva Expands NeuPro NPU Portfolio into AI Infrastructure appeared first on Embedded.