Ceva NeuPro NPU IP Portfolio

The Ceva-NeuPro portfolio gives your team a scalable family of licensable NPU IP, built on shared base technology and supported by a unified software environment. You can match each product to its workload, power and performance requirements without adopting a different architecture for every device category. 

Scale AI capability without expanding power, cost or development complexity 

AI systems increasingly combine continuous sensing, real-time perception, transformer processing, multimodal models and agentic AI workloads. Each workload places different demands on compute, memory, bandwidth, power and silicon area. 

The next-generation NeuPro portfolio gives semiconductor designers four purpose-built NPU product lines built on a common architectural and software foundation. 

With NeuPro, customers can: 

• Increase AI capability within an existing system envelope 

• Deliver equivalent capability using less power and silicon 

• Create performance headroom for future models 

• Match processing resources to each workload 

• Scale across product tiers while preserving software investment 

Scale AI inference from embedded devices to infrastructure 

Build differentiated AI products with up to 2X higher AI throughput per mm2, 50% improvement in energy efficiency*, and the right NPU architecture for every inference workload, all supported by a common toolchain.

One portfolio. Four workload-optimized NPU product lines. 

Select the inference engine that matches each product’s workload, power, memory and performance requirements. NeuPro spans always-on sensing and mainstream perception through advanced edge AI, cloud inference and neural processing within infrastructure silicon. 

* Improvement in energy efficiency versus the first generation

Bring local intelligence to power- and cost-sensitive products 

NeuPro Nano brings always-on sensing and entry-level perception to products with constrained power, silicon area and cost budgets. It enables semiconductor designers to add efficient local inference to IoT, industrial and automotive systems without overprovisioning the AI subsystem. 

Designed for: 

• Smart sensors and sensing automation 

• Predictive maintenance 

• Entry-level machine vision 

• Safety-oriented automotive and industrial applications 

Key capabilities: 

• CNN and RNN-LSTM processing 

• Tiny transformer support 

• Mixed-precision processing 

• Weight compression and acceleration 

Customer Case Studies: 

Bring transformer-based perception to mainstream devices  

NeuPro Prime delivers efficient CNN and transformer inference for mainstream vision- and audio-centric products. It enables customers to support more capable perception workloads while matching compute resources to the application’s performance, power and area requirements. 

Designed for: 

• Mainstream computer vision 

• Vision transformers 

• Vision- and audio-centric consumer products 

• Industrial perception applications 

Key capabilities: 

• CNN and transformer processing 

• Dynamic group quantization 

• Weight compression 

• Unstructured pruning 

• Data-sparsity acceleration 

Customer Case Studies:

Run multimodal, agentic and Physical AI workloads at the edge  

NeuPro Max delivers scalable processing for advanced perception, multimodal models, generative AI and agentic AI at the edge. 

It helps customers run more sophisticated models locally, supporting responsive AI experiences and complex physical-world applications. 

Designed for: 

• Multistream, high-resolution perception 

• Robotics and Physical AI 

• On-device assistants 

• Multimodal language models 

• Mixture-of-Experts architectures 

• Diffusion models 

Key capabilities: 

• Multicore scaling 

• Vector AI processing 

• Mixed-precision processing  

• Multimodal LLM and Mixture-of-Experts support 

• LLM decode acceleration 

Customer Case Studies: 

Bring programmable neural processing closer to infrastructure data  

NeuPro Ultra extends the next-generation NeuPro architecture into cloud Inference and AI infrastructure. 

NeuPro Ultra extends the next-generation NeuPro architecture to high-performance compute for AI infrastructure and inference.

By applying neural networks closer to where data is stored and moved, NeuPro Ultra can help reduce unnecessary data transfers, relieve system bottlenecks and improve infrastructure efficiency. 

Designed for: 

• Cloud AI inference 

• Edge-server inference 

• AI infrastructure silicon 

• Large multimodal and generative AI models 

Key capabilities: 

• Multicore neural processing 

• High-bandwidth memory support 

• Dynamic shapes 

• Paged KV-cache management 

• Pre-quantized LLM checkpoint support 

• Integration with infrastructure-serving frameworks 

Customer Case Study: 

Move between product tiers without starting again 

Every AI workload creates a different balance of compute, memory, bandwidth, power and silicon requirements. 

NeuPro provides four differentiated NPU product lines built on a common architectural foundation. This allows customers to right-size each design without adopting a separate NPU platform for every product category. 

Right-size every workload
Match compute, memory and processing capabilities to the application rather than overprovisioning the AI subsystem. 

Preserve software investment
Reuse development flows, tools and expertise across multiple NeuPro-based products. 

Reduce platform fragmentation
Build a broader AI product portfolio without integrating unrelated NPU architectures and software environments. 

Prepare for evolving AI models
Support established CNN workloads alongside transformers, multimodal models, Mixture-of-Experts architectures, generative AI and agentic AI. 

A unified software environment across the portfolio 

Develop once. Scale across the NeuPro portfolio. 

A common SDK and development flow helps customers evaluate models, optimize performance, compile applications and deploy inference across NeuPro product lines. 

Development teams can scale from constrained devices to advanced edge AI and multicore infrastructure systems while preserving workflows, tools and software investment. 

Evaluate
Analyze customer models and Ceva Model Zoo content using model-evaluation and performance-analysis tools. 

Optimize
Apply quantization and model optimization through a PyTorch-based frontend. 

Compile
Partition, compile and package models using the NeuPro graph-compilation flow. 

Deploy
Run CNN, transformer and LLM inference through the CevaRT runtime environment. 

Scale
Extend execution across NeuPro product lines and multicore configurations as application requirements grow.