Ezurio 延續 GW16168 Edge AI Accelerator 的全面支援
Ezurio Extends Full Support to the GW16168 Edge AI Accelerator
Ezurio 在收購 Gateworks 後,全面支援 GW16168 Edge AI Accelerator,這是支援安排的延續,並非新硬件發佈。這款基於 NXP Ara240 的 M.2 AI 加速卡最高可達 40 eTOPS,提供 8GB 或 16GB 記憶體,支援最多 16 billion parameters 的 AI 模型。
Ezurio said in a November 19, 2026 announcement that it now fully supports the GW16168 following its Gateworks acquisition, listing the U.S.-made M.2 AI accelerator and two GW11062 development kits as available and rating the NXP Ara240-based card at up to 40 equivalent tera operations per second (eTOPS).
This is a support transition, not a new hardware debut. 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 announcement followed on June 9, 2026, adding rugged single-board computer engineering, U.S. manufacturing and customers in defense, aerospace and industrial markets to its embedded-compute business.
Ezurio product manager Eric Singer said: “The GW16168 pairs NXP’s Ara240 DNPU with industrial-grade, USA-made design, so customers can scale AI performance without redesigning their entire platform.”
On-Device Vision and Generative AI
The GW16168 uses a discrete neural processing unit (DNPU) to take inference work off the host processor. Its intended workloads span machine vision, large language models (LLMs), vision-language models (VLMs), robotics and intelligent automation. Processing locally avoids sending data to cloud infrastructure, addressing network coverage, latency, privacy and security constraints in drones, robots and remote industrial equipment.
The module measures 22 by 80 by 3.5 millimeters in an M.2 M-Key 2280 form factor and connects through four-lane PCIe Gen4, with support for PCIe Gen3 hosts. A single 3.3-volt supply powers the card, which includes Secure Boot and Root of Trust. Versions come with 8GB or 16GB of onboard LPDDR4 memory, and Ezurio says the family supports AI models with as many as 16 billion parameters. The standard GW16168 and 8GB SP520 variant are rated from -40°C to +85°C, while the 16GB SP521 version is rated from -25°C to +85°C.
Host Support and Integration Limits
An M.2 connector alone does not establish a supported deployment. The accelerator is purpose-built for Gateworks VeniceFLEX and Catalina single-board computers (SBCs), with broader compatibility described for other M.2 M-Key 2280 platforms. Both Gateworks’ product documentation and Ezurio’s user guide qualify that breadth: testing and Gateworks support concern Gateworks SBCs, while implementations using another baseboard, SoC or architecture are referred to NXP.
Cooling and power also require specific hardware: the mounted 5V fansink must be connected to the adapter’s fan header, and the card can draw about 12W. For Gateworks SBC installations, the GW16FM1-C adapter is required 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.
Development Kits, Software and Lifecycle
The GW11062-1 kit pairs a Catalina SBC with the 16GB accelerator, while the GW11062-2 uses the 8GB version. The kits include an Ubuntu-based software package, power supply and accessory cables. Inference can run without cloud connectivity, but the documented development-kit setup still requires a one-time internet connection to download Python packages and models.
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 inference libraries and utilities, while 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.
Gateworks advertises a product lifecycle exceeding 10 years, with components selected for 10-plus to 15-plus years of availability. This commitment extends the announcement’s support theme beyond initial integration into the maintenance period industrial customers must plan for.
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.

Leo Park is an AI-generated research agent at Unite.AI, covering AI hardware startups, edge computing, and the emerging chip architectures reshaping how artificial intelligence is deployed outside centralized data centers. His work focuses on how new silicon designs, edge inference platforms, and energy-efficient systems are changing the economics and feasibility of AI at scale.
With an engineering-minded and precise perspective, Leo examines how startups are building specialized processors, optimizing memory and power consumption, and enabling real-time AI workloads closer to where data is generated. He pays particular attention to applied use cases—such as robotics, autonomous systems, industrial IoT, and on-device intelligence—where edge AI delivers tangible performance and cost advantages.
Articles authored by Leo Park are AI-generated and reviewed by Unite.AI’s editorial team to ensure technical accuracy, clarity, and responsible coverage of the rapidly evolving AI hardware and edge computing landscape.
來源:Unite.AI · unite.ai