As industrial IoT and edge computing continue to evolve rapidly, embedded systems increasingly require localized, high-efficiency AI inference capabilities. To address this demand, Gateworks has introduced the GW16168 M.2 AI accelerator card, designed to deliver dedicated neural network processing for complex vision algorithms and large language model (LLM) workloads at the edge. By offloading AI inference tasks from the host processor, the module helps eliminate performance bottlenecks commonly encountered in edge deployments.
Gateworks GW16168 M.2 AI accelerator
Core Architecture: NXP Ara240 Discrete Neural Processing Unit (DNPU)
At the heart of the GW16168 is the NXP Ara240 discrete neural processing unit (DNPU). This high-performance AI accelerator is purpose-built to handle compute-intensive neural network workloads while freeing the host CPU for system control and other application tasks. This architecture is particularly valuable in embedded environments where CPU resources are limited but real-time inference is required.
The module adopts a standard M.2 M-Key 2280 form factor and communicates with the host system via a PCIe Gen4 x4 interface, while remaining backward compatible with PCIe Gen3 hosts. The high-bandwidth interface ensures efficient data transfer between the accelerator and the host processor, enabling the module to fully leverage its AI inference capability of up to 40 eTOPS.
Edge AI at Scale: Supporting Models Up to 30 Billion Parameters
Compared with many conventional edge AI accelerators, the GW16168 stands out due to its onboard memory capacity and ability to support significantly larger models.
The card integrates 16GB of LPDDR4 memory, allowing neural network workloads to run independently of the host system memory. This architecture improves system stability while enabling larger and more complex models to be deployed directly at the edge.
Key advantages include:
- Large model capability – Supports models with up to 30 billion parameters (30B) when using INT4 quantization.
- Framework compatibility – Through the NXP Ara SDK, engineers can easily convert and optimize pretrained models from TensorFlow, PyTorch, and ONNX, enabling seamless migration from cloud training environments to industrial edge deployments.
This capability makes the module well-suited for applications such as advanced machine vision, industrial inspection, intelligent transportation systems, and predictive maintenance platforms.
NXP Applications Processor and Ara-DNPU Connection
Industrial-Grade Reliability and Thermal Efficiency
The GW16168 is engineered for demanding industrial environments. The accelerator typically consumes approximately 6.6W, offering a favorable performance-per-watt ratio that makes it well suited for fanless embedded systems.
The product is designed, tested, and assembled in the United States, ensuring supply-chain transparency and high manufacturing quality standards.
Additional features include integrated secure boot and hardware root of trust, providing strong device integrity protection for edge AI systems handling sensitive operational data.
| Category | Specification |
|---|---|
| Core NPU | NXP Ara240 Discrete Neural Processing Unit (DNPU) |
| AI Performance | Up to 40 eTOPS |
| Onboard Memory | 16GB LPDDR4 |
| Host Interface | PCIe Gen4 x4 (backward compatible with PCIe Gen3) |
| Form Factor | M.2 M-Key 2280 |
| Model Support | Up to 30B parameters (INT4 quantization) |
| Security Features | Secure Boot and Hardware Root of Trust |
| Operating Temperature | -40°C to +85°C (Industrial grade) |
| Typical Power Consumption | ~6.6W |
Machine learning deployment flow
Deployment Flexibility for Embedded Systems
For system integrators and developers, the GW16168 provides significant deployment flexibility. The accelerator can be integrated directly into Gateworks embedded platforms such as the VeniceFLEX and Catalina single-board computers, or installed in any embedded system equipped with an M.2 M-Key slot.
Beyond raw compute capability, the module delivers end-to-end system robustness. Secure Boot ensures the integrity of software running on edge devices, while the industrial temperature range of –40°C to +85°C enables reliable operation in demanding environments—from outdoor traffic monitoring systems to automation equipment deployed in harsh industrial facilities.
The GW16168 AI accelerator card and its accompanying development kit are expected to begin shipping in late May. Once available, customers will be able to purchase the product through major global distribution channels including DigiKey, Braemac, RoundSolutions, and Avnet.
