Why Edge Computing Matters

Unlike centralized cloud models (AWS, Azure, GCP), edge computing runs workloads near the data source inside a plant, camera, or device.

Key advantages:

• Low latency: milliseconds instead of cloud round-trip delays.
• Resilience: workloads keep running even with poor connectivity.
• Cost efficiency: avoids heavy cloud GPU usage.
• Security & compliance: sensitive data stays local.

How to classify edge computing types?

IIoT scenarios such as image processing, text analysis, and cyber security demand intense computational power. In these cases, Neural Networks are utilized, making Edge Computing and analytics essential. Recently, we worked on projects requiring Edge Computing devices capable of running AI models. We explored available options for edge devices and compared five major competitors in the current market based on our findings.

Here are the contexts for deploying edge devices in our two projects:

1. Electroplating Plant: Develop a Neural Network model for Process Modeling and Control to determine the optimal control settings for the plant.
2. OEM Manufacturing Plant: Automatically inspect defects for Quality Control using cameras.

We analyzed the pros and cons of various market options and summarized our findings in the table below.

‍

These options were analyzed not only for features but also for compatibility with manufacturing assets, supported OS, and so on.

We decided to zero in on NVIDIA Jetson for those particular implementations, mainly because the Nano is built on a 128 Core Maxwell GPU which is used on GeForce graphic cards. It also supports a lot of ML frameworks with a higher level of performance compared to others.

Conclusion: 

In our evaluation of edge computing devices for IIoT applications involving image processing, text analysis, and cybersecurity, NVIDIA Jetson emerged as the preferred choice due to its robust performance and compatibility with a variety of ML frameworks. The NVIDIA Jetson Nano’s 128 Core Maxwell GPU offers significant computational power, making it well-suited for the Neural Network models required in our projects.

We rejected the others because:

1. Google Coral’s TPUs: Built for high volume, low precision processing and can only work with TensorFlow.
2. Intel Up2 with Compute Stick 2 or Compute Stick 2 with any Ubuntu/Raspberry Pi: Provides SDK for Computer Vision (Open Vino), which is not supported in Windows.

Google Coral is the latest in the market and sounds promising. We’ll certainly be keeping an eye on it and might even recommend it to our clients if there’s a feature fit