Chips & Compute

Ezurio Extends Full Support to GW16168 Edge AI Accelerator

Ezurio now fully supports the GW16168, the NXP Ara240-based M.2 edge AI accelerator rated at 40 eTOPS, following its June 9, 2026 acquisition of Gateworks. The card runs LLMs of up to 16 billion parameters on-device.

Ezurio Extends Full Support to the GW16168 Edge AI Accelerator
Ezurio Extends Full Support to the GW16168 Edge AI AcceleratorPexels / pixabay
By Elena Vasquez4 min read

Why it matters

  • Ezurio announced full support for the GW16168 on November 19, 2026, following its June 9, 2026 acquisition of Gateworks.
  • The GW16168 is rated at up to 40 eTOPS and based on NXP's Ara240 DNPU, with 8GB or 16GB LPDDR4 variants supporting models up to 16 billion parameters.
  • The module ships in M.2 M-Key 2280 form factor with PCIe Gen4, draws about 12W, and is rated from -40°C to +85°C (8GB SP520) or -25°C to +85°C (16GB SP521).
  • Two GW11062 development kits are available, pairing a Catalina SBC with the 16GB or 8GB accelerator plus Ubuntu-based software.
  • Gateworks advertises a product lifecycle exceeding 10 years, with components selected for 10-plus to 15-plus years of availability.

Ezurio announced on November 19, 2026 that it now fully supports the GW16168, a U.S.-made M.2 edge AI accelerator rated at up to 40 equivalent tera operations per second (eTOPS) and built on NXP's Ara240 discrete neural processing unit.

The announcement is a support transition, not a new hardware launch. Gateworks introduced the GW16168 on April 7, 2026, after unveiling it with NXP at Embedded World 2026, and listed it as available directly and through distributors. Ezurio's acquisition of Gateworks followed on June 9, 2026. That deal added rugged single-board computer engineering, U.S. manufacturing, and customers in defense, aerospace and industrial markets to Ezurio's embedded-compute business.

Ezurio product manager Eric Singer framed the module's value in a statement: "The GW16168 pairs NXP's Ara240 DNPU with industrial-grade, USA-made design, so customers can scale AI performance without redesigning their entire platform."

Alongside the card itself, Ezurio lists two GW11062 development kits as available — one built on the Catalina single-board computer with the 16GB accelerator, the other with the 8GB version.

What does the accelerator actually do?

The GW16168 uses its discrete neural processing unit to take inference work off the host processor. Target workloads span machine vision, large language models (LLMs), vision-language models (VLMs), robotics and intelligent automation.

Processing locally matters for a specific set of constraints. The card avoids sending data to cloud infrastructure, which addresses network coverage, latency, privacy and security limits in drones, robots and remote industrial equipment. In practice, that's the difference between a vision system that keeps working in a coverage dead zone and one that doesn't.

The hardware specs are concrete:

  • Form factor: M.2 M-Key 2280, measuring 22 by 80 by 3.5 millimeters
  • Interface: four-lane PCIe Gen4, with support for PCIe Gen3 hosts
  • Power: a single 3.3-volt supply, with the card drawing about 12W
  • Memory: 8GB or 16GB of onboard LPDDR4
  • Security: Secure Boot and Root of Trust
  • Model size: Ezurio says the family supports AI models with as many as 16 billion parameters

Thermal ratings differ by variant. The standard GW16168 and the 8GB SP520 are rated from -40°C to +85°C. The 16GB SP521 is rated from -25°C to +85°C.

What are the integration limits?

An M.2 connector alone does not establish a supported deployment, and both companies are explicit about that. The accelerator is purpose-built for Gateworks VeniceFLEX and Catalina single-board computers, with broader compatibility described for other M.2 M-Key 2280 platforms.

The fine print narrows that breadth considerably. Gateworks' product documentation and Ezurio's user guide both qualify that testing and Gateworks support concern Gateworks SBCs only. Implementations using another baseboard, SoC or architecture are referred to NXP.

Cooling and power require specific hardware too:

  • The mounted 5V fansink must be connected to the adapter's fan header
  • The card can draw about 12W
  • Gateworks SBC installations require the GW16FM1-C adapter, because the basic adapter lacks sufficient power and a fan header

A standard connector therefore describes the physical interface, not the complete thermal, power and support design.

What's in the development kits?

The GW11062-1 kit pairs a Catalina SBC with the 16GB accelerator. The GW11062-2 uses the 8GB version. Both kits ship with an Ubuntu-based software package, a power supply and accessory cables.

There's one networking caveat. Inference can run without cloud connectivity once everything is set up, but the documented development-kit setup still requires a one-time internet connection to download Python packages and models.

How does the software stack work?

The software path supports TensorFlow, PyTorch and ONNX models through NXP's Ara SDK. Its extensible compiler covers convolutional neural networks, vision transformers, LLMs and VLMs, with support for INT4, INT8 and MXINT8 quantization data types where applicable.

NXP's runtime supplies the inference libraries and utilities. Its eIQ AAF Connector exposes locally hosted Ara240 LLM and VLM backends through an OpenAI-compatible REST API, giving client applications a chat-completions-style interface without moving inference off the device. That detail matters for developers: existing OpenAI-style client code can target an on-device model with minimal changes.

How long will the hardware last?

Gateworks advertises a product lifecycle exceeding 10 years, with components selected for 10-plus to 15-plus years of availability. That commitment extends the announcement's support theme beyond initial integration into the maintenance period industrial customers must plan for.

This is where the acquisition story meets the engineering one. Industrial deployments don't get refreshed like consumer gadgets; a module that stays orderable and supported for a decade plus is a procurement requirement, not a bonus. Ezurio's backing, with its U.S. manufacturing footprint and defense and aerospace customer base, is the signal the November 19 announcement is designed to send.

For embedded designers, the practical change is a clearer support path around hardware that was already available. The GW16168 offers a modular route to local vision and generative-AI inference, but production deployment still depends on host-specific support, adequate cooling and power, and the required software setup. Those integration details, alongside the lifecycle commitment, will determine how easily the module moves from an evaluation kit into a maintained industrial system.

Source: unite.ai

Share this article:

More from Elena Vasquez

Elena Vasquez

Show full bio

Market editor covering technology and product news in artificial intelligence and machine learning at AI In Context.

795 articles

Related articles

  1. Alibaba unveils Zhenwu V900 AI chip, targets 20 GW cloud buildout by 2032
  2. Naver Lands $1.5 Billion AI Computing Deal With US Inference Firm
  3. Argentum AI, Kyndryl announce $630M Boulder data center deal
  4. Expedera, Cadence report validated pre-silicon edge AI integration
  5. Huawei and Alibaba Chip Blitz Arms Xi With Leverage Ahead of Trump Summit

« Previous articleNext article »