Enabling Efficient Edge AI with DEEPX NPUs and the Avnet Silica Ecosystem
Artificial intelligence is rapidly moving from the cloud to the edge. From autonomous robots navigating complex environments to smart cameras inspecting production lines, many applications now require AI to run directly on embedded devices, where power, latency, and thermal constraints are critical. However, deploying high-performance AI at-the-edge has traditionally required large Graphics Processing Units (GPUs) or power-hungry processors, making it difficult to build compact, energy-efficient systems.
DEEPX is rewriting the rule book. Its family of ultra-efficient Neural Processing Units (NPUs) enables powerful AI inference directly on edge devices while consuming only a fraction of the power of conventional AI accelerators. Together with Avnet Silica’s hardware platforms, engineering expertise, and partner ecosystem, developers and system designers can rapidly build and deploy scalable edge AI solutions.
A brief overview of DEEPX NPUs
DEEPX has developed a range of dedicated AI accelerators designed specifically for on-device inference at the edge. These NPUs are optimised to deliver high performance while maintaining extremely low power consumption and minimal thermal output.
For example, the DX-M1 NPU can perform advanced AI tasks such as video analytics, object detection, and sensor data processing while consuming only a few watts of power. To visually demonstrate the efficiency of the DEEPX NPU is the so-called ‘butter-proof’ benchmark test (Figure 1).
In this test, a piece of butter is placed directly on the NPU while it runs demanding AI workloads such as video analysis. Because butter melts at roughly 30 °C to 36 °C, it provides a simple visual indicator of how much heat the chip generates. During the demonstration, the butter does not melt, illustrating the DEEPX chip's extremely low heat output and power consumption.
The test highlights one of the key advantages of the DEEPX architecture: high AI performance without the need for large heatsinks or active cooling, making it ideal for embedded and edge systems where space and thermal budgets are limited.
Figure 1: Benchmarking the DEEPX NPU using the ‘butter-proof’ test illustrates its extremely low heat output and power consumptionBy processing AI workloads locally, edge devices gain several advantages:
- Ultra-low latency for real-time decision making
- Reduced bandwidth usage by avoiding cloud data transfers
- Improved privacy and security by keeping data local
- Higher reliability even in environments with limited connectivity
This architecture opens the door to a new generation of intelligent edge devices across industries, including robotics, manufacturing, medical, security, and smart infrastructure.
Edge AI in action
Since the launch of engineering samples in 2023, DEEPX has experienced rapid growth in market interest, engaging with over 350 global companies for testing, and now has over 50 customer projects in full production.
Hyundai
Working with Hyundai in Korea, the company replaced the GPU - needed to run two AI models concurrently - with its DX-M1 connected to a general-purpose ARM processor. In this smart mobility application, the commercial delivery robots are used in urban areas for last-mile food and package deliveries.
The robot runs facial recognition models on its camera (Figure 2). Upon delivery, it matches the photo of the customer taken from the smartphone used at the time of purchase. The robot can then automatically verify the recipient and open the hatch of the appropriate storage compartment to release the food or goods.
In parallel to the facial recognition models, another camera runs a positional awareness model to determine the position of the recipient’s feet – you do not want the robot to run over their toes. Using the DX-M1 device running at less than 5 W instead of a 20 W, 30 W or 40 W GPU-based application reduces the number of charge cycles significantly.
Figure 2: Commercial delivery robot using the DEEPX DX-M1 – finished verification and ready for mass production
Ultralytics
DEEPX has also supported various types of factory automation applications. Working in close collaboration with Ultralytics, the company behind the YOLO object detection models, two particular use cases stand out.
The first is for defect detection. As shown on the right in Figure 3, AI-enabled cameras are used to detect scratches and dents on the product’s casing. It also runs object character recognition (OCR) on the letters on the barcode to identify the defective product. The right-hand image illustrates how Ultralytics used the DX-M1 in combination with an x86 host, to count anything from tins to bottles. Neither of these application types is unique, but both solutions can run complex models on multiple cameras at below 5 W and at a lower cost.
Figure 3: Automated quality control on a fast-moving production lineAccelerating Edge AI with the Avnet Silica ecosystem
While powerful silicon is essential, bringing Edge AI products to market requires far more than a processor. This is where Avnet Silica plays a critical role. Working closely with DEEPX, Avnet Silica provides a complete ecosystem that helps designers move from concept to production faster.
This ecosystem includes:
- Processor partnerships: Compatibility with leading application processors, using the PCIe interface, from vendors such as AMD, NXP, Renesas, TRIA, Microchip, STMicroelectronics, and Qualcomm, enables flexible system architectures.
- Hardware platforms: Industrial PCs, embedded modules, and edge computing platforms that integrate DEEPX NPUs like Advantech, avalue, ASRock, ASUS IoT, Engicam, GigaIPC, Kontron, and MSI.
- Software and OS: Avnet Silica’s technical teams help customers design, integrate, and optimise AI systems tailored to their applications with suppliers including Ubuntu Linux, Yocto Linux, and Microsoft Windows 11 IOT.
- AI ecosystem partners: From robotics platforms to industrial automation systems, the broader partner network helps accelerate real-world deployment.
Together, DEEPX and Avnet Silica provide a pathway for companies to adopt edge AI without building the entire infrastructure themselves.
Enabling developers with the DEEPX SDK
A powerful chip also requires accessible tools for developers. The DEEPX software development kit (SDK) provides the environment needed to deploy Edge AI models efficiently on DEEPX hardware (Figure 4).
The SDK supports:
- AI model optimisation and compilation
- Deployment of vision AI and deep learning models
- Integration with common AI frameworks
- Performance tuning for edge environments
Developers can adapt existing AI models for efficient inference on DEEPX NPUs, enabling rapid prototyping and production deployment. Combined with Avnet Silica’s hardware platforms and engineering expertise, the SDK helps transform AI concepts into scalable edge solutions.
Figure 4: DEEPX SDK overview with AI model compile and runtime environmentsTurning Edge AI innovation into reality
Edge AI is transforming industries, from autonomous robotics to smart manufacturing, by enabling devices to process data and make real-time decisions. With its highly efficient NPUs, DEEPX enables the execution of powerful AI workloads directly on embedded systems without the power and thermal limitations of conventional accelerators.
Through its collaboration with DEEPX, Avnet Silica provides the hardware platforms, ecosystem partnerships, and engineering support needed to bring these solutions to market faster. The end goal is to help companies unlock the full potential of Edge AI.
Ready to bring AI to your edge devices? Avnet Silica can help you explore the DEEPX platform, integrate the technology into your designs, and accelerate your transition from prototype to production. Contact Avnet Silica to learn how the DEEPX ecosystem can power your next generation of Edge AI solutions.
Frequently asked DEEPX/NPU questions
| Question | Answer |
|---|---|
| What is the difference between TOPS and equivalent TOPS? |
TOPS (Trillion Operations Per Second) is a theoretical measure of a processor's peak mathematical capacity. However, raw TOPS can be misleading if the hardware suffers from memory bottlenecks or inefficient data handling. eTOPS (equivalent TOPS) measures effective throughput—the actual work delivered on real-world AI models. DEEPX achieves high eTOPS by using architectural innovations like weight compression, sparsity support (skipping zero-value calculations), and near-memory computing. This allows a DEEPX NPU with lower power consumption to match or exceed the frames-per-second (FPS) of a high-wattage GPU with a higher "paper" TOPS rating. Essentially, eTOPS represents "useful" AI performance rather than just raw, idle cycles. |
| Why are FPS/TOP and FPS/W important when comparing GPUs and NPUs? |
When comparing AI processors, especially GPUs and dedicated NPUs used in vision systems, two additional metrics are very useful:
|
| Why are dedicated NPUs becoming more popular for edge AI? | While GPUs are extremely powerful and flexible, they were originally designed for graphics workloads and large-scale compute environments. Dedicated NPUs are designed specifically for neural network inference, which allows them to deliver higher performance per watt, lower thermal output, smaller silicon footprints, and optimised execution of AI models. |
| How easy is it to port existing AI models to DEEPX NPUs? | Porting existing AI models to DEEPX NPUs is designed to be relatively straightforward, especially if the model was originally trained using common deep learning frameworks. The key reason is that DEEPX uses the ONNX (Open Neural Network Exchange) format as its primary model input. ONNX serves as a universal model representation, enabling neural networks trained in frameworks such as PyTorch, TensorFlow, or scikit-learn to be exported to a common format for deployment across different hardware platforms. |
On Demand Webinar: Deploying Edge AI at GPU-Level Performance with DEEPX
Join Avnet Silica and DEEPX for an exclusive one-hour webinar introducing cutting-edge AI inference acceleration technology now available in Europe. As the demand for edge AI continues to grow, traditional GPU solutions often fall short in power efficiency and thermal constraints. DEEPX addresses this critical gap with purpose-built edge-first AI processors that deliver exceptional performance per watt while maintaining high-precision inference.
Over 240 attendees joined us for the live webinar, making it one of our most-attended online events to date. Catch up below!
