Ceva’s next NeuPro generation opens new possibilities across AI

Ceva is broadening the possibilities for AI silicon with a new generation of its NeuPro neural processing unit portfolio, bringing a common architecture and unified software environment to applications spanning sensors, intelligent devices and cloud infrastructure.

Announced on October 28, the portfolio reflects an opportunity: helping semiconductor companies deliver more capable AI within practical power, memory and silicon constraints. Ceva says the architecture provides up to twice the AI throughput per square millimeter and a 50% improvement in energy efficiency compared with its first generation.

These company claims await detailed customer benchmarks, but the direction is encouraging. Greater efficiency could give developers more freedom to introduce sophisticated AI features while managing product size, energy consumption and cost.

A platform designed to grow

The portfolio comprises four NPUs, each addressing a different level of inference demand.

NeuPro Nano targets continuous sensing and entry level perception in constrained devices. NeuPro Prime supports mainstream convolutional neural networks and transformer workloads. NeuPro Max scales toward advanced edge applications, including robotics, multimodal models, device assistants and agentic AI.

NeuPro Ultra extends the platform into AI infrastructure, supporting complex language model inference and integration into intelligent memory, connectivity and data processing silicon.

Together, these options give developers a broader foundation for building products around different performance requirements. Companies can select an appropriate engine, combine several within one system on chip, or move between portfolio members as their roadmaps evolve.

A unified software development kit supports that flexibility. Reusing tools and development approaches could help engineering teams preserve their software investment, reduce integration complexity and bring successive products to market more efficiently.

Making demanding models more practical

The architecture addresses more than raw processing capacity. Ceva has added capabilities intended to improve how models use memory, move data and occupy computing resources.

Support for FP4, FP8 and mixed precision gives developers additional ways to balance computational efficiency and model requirements. Higher memory bandwidth, dynamic shapes and paged key value caching address challenges associated with transformer models and longer contexts.

Support for prequantized models could simplify deployment by allowing customers to work with published checkpoints without an additional quantization and calibration stage. Compression, pruning and optimized tensor reshaping offer further opportunities to improve effective throughput.

The architecture also supports Mixture of Experts and linear attention. That breadth is valuable for developers planning products in a market where model designs continue to change.

The opportunity is practical: helping teams translate advances in AI research into deployable features, with more choices for meeting performance, memory and energy budgets.

Why it matters

Ceva’s expansion comes as AI development increasingly requires coordination across devices, edge systems and cloud infrastructure. Different tasks benefit from different execution environments, making adaptable processing platforms useful.

For device makers, efficient inference could enable richer local perception, more responsive assistants and increasingly capable machines. Processing selected workloads nearby can also reduce the need to send every task to remote infrastructure.

For infrastructure designers, NeuPro Ultra creates another route toward specialized acceleration. Its proposed integration into memory and connectivity subsystems could help address data movement bottlenecks closer to their source, alongside its role in running models.

Ceva’s licensing approach is significant. Semiconductor companies can draw on an existing NPU architecture and software foundation rather than developing every component internally. That could widen participation in AI silicon development and free engineering resources for product differentiation.

The company says select customers already have early access, with general availability planned for the second quarter of 2027. Those engagements provide a path toward refining the technology around practical deployment needs.

Bottom line: The portfolio presents a solid proposition: a consistent development foundation that expands from small sensing tasks to demanding infrastructure applications. CEVA NeuPro will help companies build useful AI products with greater efficiency, flexibility and confidence across their future semiconductor roadmaps and product generations.

Also Read:

Ceva’s Next-Generation NeuPro Brings a Unified AI Platform from Sensors to Infrastructure

UWB Without the Integration Headache: Ceva and LG Bring Precision to More Chips

2026 Ceva Edge Symposium: Building Intelligence Beyond the Cloud