text.skipToContent text.skipToNavigation

Truck-Icon.svg Free shipping within the continental US over $50. Conditions apply

  1.  
  2. Resources
  3. Featured Products
  4. NXP — Ara-2 Discrete Neural Processing Units (DNPUs)

NXP — Ara-2 Discrete Neural Processing Units (DNPUs)

 

SKINNY-NXP-Kinara-DNPUs

 

NXP

NXP Ara Family of Discrete NPUs for Edge AI Acceleration

Drive what’s next in Edge AI

NXP’s Ara family of discrete neural processing units, part of NXP’s edge AI portfolio, empowers designers to push the boundaries of intelligent systems. By offloading intensive AI inference from host processors, Ara NPUs deliver low-latency, power-efficient AI execution while providing the flexibility to scale as models and applications evolve.

With Ara-2 accelerating real-time generative AI and Ara-1 serving established deployments, the portfolio allows system designers to future-proof their AI strategies, improving privacy, responsiveness, and overall system efficiency.

Ara-2 Discrete NPU for Real-Time Generative AI

Ara-2 is NXP’s latest discrete neural processing unit, optimized for real-time generative AI and modern transformer-based workloads at the edge. Built to handle the compute and memory demands of large language models and multimodal AI, Ara-2 enables on-device execution with low latency, reduced operational costs and enhanced data privacy.

Ara240 is the first product launching in this family, and currently in preproduction. Its balanced architecture combines high compute density, large on-chip memory and high off-chip bandwidth to efficiently execute large and complex models across embedded and AI-enabled compute platforms.

Ara240 Key Features

  • AI Model Framework Support
    • TensorFlow
    • PyTorch
    • ONNX
  • AI Model Architectures Supported
    • Convolutional neural networks
    • Transformer models
    • Large language models
    • Multimodal language models
    • Vision language models
    • Vision language actions
  • Performance
    • Up to 40 equivalent tera operations per second (eTOPS)
  • Security
    • Secure boot
    • Root-of-trust processor
  • Memory Interface
    • Up to 16 GB LPDDR4
  • Operating System Support
    • Linux
    • Windows
  • Host Interfaces
    • PCIe Gen4 x4
    • USB 3.2 Gen 2
  • Package
    • 17 mm x 17 mm flip-chip ball grid array

Applications

  • AI assistance
  • Factory automation
  • Smart retail
  • Home entertainment
  • Energy management systems
  • Industrial human machine interfaces
  • Patient monitors


NXP — Ara-2 Discrete NPU

Block Diagrams
NXP Applications Processor and Ara DNPU Connection
Integration of Ara DNPU and Applications Processors
Machine Learning Deployment Flow

Edge AI Acceleration within the NXP Portfolio

Following the completion of NXP’s acquisition of Kinara, Ara discrete NPUs are now part of NXP’s broader intelligent edge portfolio. Ara devices pair naturally with NXP application processors, enabling system-level AI acceleration where NPUs handle inference while the host processor manages control, preprocessing and postprocessing.

The Ara software stack is being integrated into NXP’s eIQ AI and ML software environment, simplifying model optimization, deployment and scaling across embedded platforms.

NXP technologies enable flexible, scalable AI systems that support everything from traditional time series anomaly detection or vision inference, to advanced, multimodal generative AI at the edge.

Hardware −
Ara240 16GB M.2 Module
The Ara240 16GB M.2 Module is an AI accelerator module based on the Ara240 AI processor. It can provide up to 40 eTOPs with flexible precision. It uses the M.2 M-Key and supports a maximum 4 lanes PCI Express 4.0 in x4 configuration. Designed with the commonly used NGFF 2280 form factor, it can be easily used as a companion processing unit on any NXP i.MX or S32 platform with M.2 M-Key PCI Express capability.
Ara-2 USB Module: eTOPS for Gen AI Workloads
The Ara-2 USB module delivers up to 40 eTOPS for high-performance, power-efficient AI inference. It supports real-time AI workloads, including generative AI, large language models (LLMs) and vision language models (VLMs), enabling compute and embedded systems to run multiple models simultaneously without performance penalties. The module can process visual and textual content, offering enhanced productivity and capabilities for edge-AI deployments.