Discrete Neural Processing Units for AI Acceleration

We provide two discrete neural processing units (DNPUs): Ara-1 is our first generation DNPU that is ideal for energy efficient AI inferencing at the edge in cameras, embedded systems and edge servers. Ara-2, capable of up to 40 eTOPS (equivalent Tera Operations Per Second), is our second generation DNPU, optimized for real-time generative AI. Available as chips or modules, the Ara-1 and Ara-2 DNPUs can be easily added to enhance embedded systems with advanced AI capabilities, including upgrading existing in-field systems. This makes them the preferred solution for emerging AI applications in vision, voice, gesture and a variety of other multimodal AI implementations.

Neural processing units (NPUs) are specialized processors architected to accelerate the execution of neural networks. The localized processing of NPUs leads to dramatically faster response times, significantly enhanced privacy and reduced latency, all of which are critical for numerous industrial and automotive applications.

Having both discrete NPUs and integrated NPUs offers flexibility. DNPUs are ideal for high-performance, low latency specialized AI tasks, while integrated NPUs can handle more general-purpose AI processing and lower-power always-on AI functions. This combination allows for a more optimized and efficient intelligent edge system architecture.

When combined with NXP’s Applications Processors, our DPNU portfolio provides customers with an end-to-end platform that delivers very efficient front-end processing and high-performance AI workload acceleration. NXP's customers will have access to a comprehensive and powerful portfolio of AI solutions, ranging from low-power TinyML applications to high-performance Generative AI implementations to being able to run multi-modal Agentic AI workflows seamlessly. While the integration with NXP Applications Processors provides an enhanced developer experience, the NXP DNPUs can also be used with non-NXP host processors.

Our DNPU products stand out by offering a programmable, flexible and efficient compute and memory architecture that enables both CNNs and leading-edge transformer-based models including visual language models (VLMs), Large Language Models (LLMs) enabling advanced Agentic-AI use-cases. Our DNPUs support dynamic quantization, enhancing performance while reducing system costs by lowering memory requirements. Furthermore, the architecture of NXP DNPUs is designed to process multiple models without incurring switch-time performance penalties.

The combined technologies will unlock a wide spectrum of edge AI applications, including:

• Multimodal generative AI

  Creating truly interactive and intelligent devices capable of understanding and responding to diverse inputs like voice, vision and text. This aligns with the growing trend of multi-modal generative AI models, which demonstrate the power of combining different data modalities for richer understanding

• Advanced vision processing

  Enabling highly sophisticated object detection, image recognition and video analytics for applications like autonomous driving, smart surveillance and industrial automation

• Voice and natural language processing

  Improving voice assistants, speech recognition and natural language understanding for more intuitive and seamless human-machine interaction

• Industrial automation

  Enabling predictive maintenance, anomaly detection and real-time process optimization, leading to increased efficiency and reduced downtime

Yes, the DNPU software enablement will be integrated into NXP eIQ AI SW development environment to create a seamless experience for our customers to use all of NXP’s Edge AI products. The combined development environment will include model libraries and model optimization tools, providing developers with more powerful tools for optimizing AI models for edge deployment.