In industrial projects, choosing an edge computer is rarely about buying the "most powerful hardware." The real challenge is finding a platform that can handle today's workloads while remaining flexible enough for future AI requirements. You don't want to overpay for compute you'll never use, but you also don't want to tear out a gateway six months later when the project scope inevitably expands.
A scalable edge computing architecture provides a practical way to address this challenge. The following discussion uses a specific platform as an example.
1. 6 TOPS Is the Starting Point — Scalable AI Computing up to 320 TOPS
This industrial edge platform is powered by the Rockchip RK3588J octa‑core industrial processor, featuring 4× Cortex‑A76 performance cores and 4× Cortex‑A55 efficiency cores, with a standard configuration of 8 GB RAM and 128 GB eMMC storage. The built‑in NPU delivers 6 TOPS of computing power, supporting mixed‑precision operations across INT4, INT8, INT16, and FP16.
What does 6 TOPS actually mean in practice? Running YOLOv8‑based object detection on a single 1080P video stream, inference latency can be kept under 50 ms depending on model size and deployment configuration. When running 4 industrial cameras simultaneously, CPU utilisation stays below 65%.
But what if the project needs to run large models? For example, equipment predictive maintenance may require a 7B‑parameter time‑series analysis model, or production line quality inspection may need to run a 35B‑parameter multimodal large model. 6 TOPS is clearly not enough.
The device's solution is to reserve 2× M.2 PCIe high‑speed interfaces on the chassis, supporting the addition of compute expansion modules. Optional R182X or LQ50 accelerator cards are available, with maximum heterogeneous computing power reaching 320 TOPS. Performance testing shows that dual LQ50 modules can simultaneously run inference workloads with two 35B‑parameter large models under the tested configuration, and the Qwen3.5 35B‑A3B model achieves a Decode TPS of 24.06. This means a single unit can serve simultaneously as a data acquisition gateway, an edge AI inference server, and a local large‑model inference platform.
This elastically expandable compute design essentially uses a combination of a general‑purpose main controller and dedicated AI accelerators to flexibly cover the full range of scenarios, from lightweight inference to complex industrial‑grade tasks.
2. Three Software Layers for Faster Industrial Application Development
Without the right software ecosystem, even powerful hardware cannot deliver practical industrial applications. The device comes pre‑installed with the NeuronEX‑Lite industrial protocol hub, co‑developed by IOTRouter and EMQ, plus the Node‑RED visual programming engine and the FUXA zero‑code configuration tool, forming a three‑layer software architecture covering protocol connectivity, visual programming, and industrial visualization.
NeuronEX‑Lite handles southbound protocol access. It is compatible with over 100 industrial protocols, including Modbus RTU/TCP, OPC UA, EtherNet/IP, Siemens S7, Omron, Allen‑Bradley, and more, enabling low‑latency industrial data acquisition and processing. Node‑RED provides drag‑and‑drop workflow orchestration, with over 50 pre‑built industrial nodes — from reading PLC data to publishing via MQTT; just connect the nodes and deploy. FUXA handles web‑based visual configuration, with a library of over 200 components supporting multi‑screen rendering, enabling zero‑code construction of data dashboards and process flow diagrams.
The practical effect of this combination is that, in many industrial scenarios, IoT application development cycles can be shortened from weeks to days. Protocol access, data processing, business logic, and visual presentation are all completed on a single EC700 device.
3. IT and OT Convergence: Southbound Manages Heterogeneous Devices, Northbound Directly Connects to Major Clouds
The most vexing problem on industrial sites is not technology itself — it is heterogeneity. Siemens PLCs, Mitsubishi inverters, Omron sensors — each speaks its own protocol. The upper‑layer platform might be Alibaba Cloud IoT, OneNET, or a customer's self‑built private cloud.
This edge platform is positioned as an IT/OT integration hub. On the southbound side, it manages heterogeneous devices through NeuronEX‑Lite; on the northbound side, it directly connects to various cloud platforms via MQTT, HTTP, TCP, UDP, WebSocket, and other protocols. In a project test environment, the southbound side simultaneously connected a Modbus RTU power meter, an OPC UA machine tool, and a Siemens S7‑1200 PLC, while the northbound side connected to the Alibaba Cloud IoT Platform — the entire configuration took less than 30 minutes.
In many industrial deployments, integrating gateway, edge computing, and visualization functions into one platform can significantly reduce hardware investment and system complexity. Fewer devices mean fewer cables, fewer support calls, and fewer failure points — which translates directly to lower total cost of ownership.
4. Industrial‑Grade Fanless Thermal Design — Reliable Operation from ‑20°C to 70°C
The environmental challenges on industrial sites far exceed those faced by ordinary commercial equipment. High temperatures, dust, and electromagnetic interference are the norm. The EC700 uses an all‑metal finned chassis with dual pure‑copper heat pipes, with the SoC directly contacting the thermal base, establishing a complete thermal path from the chip to the heat pipes to the outer chassis. The fanless design prevents dust accumulation, and the operating temperature range is ‑20°C to 70°C.
In terms of electrical safety, it uses a DC 12 V power supply with reverse‑connection protection. All interfaces are isolation‑designed: RS485 features galvanic isolation, DI uses optocoupler isolation, and DO uses relay outputs with a load capacity of 5 A @ 30 VDC / 5 A @ 250 VAC. EMC testing has passed electrostatic discharge Level 3 (±8 kV air discharge, ±4 kV contact discharge), and both surge and burst tests have reached Level 2 (±1 kV), meeting relevant industrial EMC testing standards.
Hardware watchdog plus dual software watchdog mechanisms ensure automatic reset under extreme anomalies, enabling 24/7 unattended stable operation.
5. Multi‑Network Intelligent Switching and Integrated Visual and Audio Monitoring Capability
Network stability is the lifeline of industrial IoT. The EC700 supports 4G/5G + dual Gigabit Ethernet ports + Wi‑Fi 6 multi‑mode network access, with automatic reconnection on drop and multi‑channel hot‑standby. The 4G module uses Quectel Cat.4, supporting LTE‑FDD bands B1/B3/B5/B8 and others; the 5G module supports SA n1/n28/n41/n77/n78/n79 and NSA n41/n78/n79 bands. Wi‑Fi 6 supports dual‑band 2.4 GHz + 5 GHz with a maximum throughput of 600.4 Mbps.
For visual and audio monitoring, the device provides HDMI input and output interfaces, supporting 8K video decoding and 4K encoding, and can directly connect to large screens to display production status in real time. The 3.5 mm audio jack supports device warning audio output. This integrated visual/audio design is particularly practical in production line quality inspection scenarios: industrial camera feeds are displayed in real time on the HDMI large screen, AI inference results are overlaid with annotations, and audio alerts sound on anomalies — operators can grasp the full situation without leaving their workstations.
6. Boundary‑Free Remote Operations and Maintenance — Debug PLCs from Thousands of Miles Away
The more dispersed the devices, the higher the operations and maintenance costs. The EC700 comes with IOTRouter's self‑developed free remote O&M software, supporting remote configuration, remote debugging, remote diagnostics, and remote updates. Core capabilities include virtual serial ports and virtual network interfaces, allowing engineers to debug downstream devices from their computers with the same experience as plugging a cable in on‑site.
More critically, it has built‑in NAT traversal and P2P hole‑punching technology. Even if the device is deployed behind a carrier‑grade NAT or corporate firewall, no public IP address and no router port forwarding are required to enable secure remote access. The technical principle is that the EC700 actively maintains a persistent connection to the cloud handshake server; when an O&M engineer initiates an access request, the cloud assists in establishing a point‑to‑point encrypted tunnel, reducing unnecessary cloud forwarding and improving latency and bandwidth efficiency.
7. Secure and Flexible Industrial Computing Architecture
The EC700 uses a hardware and software architecture designed for supply‑chain reliability and long‑term industrial deployment. The core processor, Rockchip RK3588J, is paired with an embedded Linux operating system, and the software stack — Node‑RED and NeuronEX‑Lite — is either open‑source or customised for industrial use. This approach provides flexibility, transparency, and the ability to adapt to specific project security requirements.
For customers with volume requirements, IOTRouter also offers OEM/ODM services, including hardware silk‑screen customisation, firmware feature customisation, and private deployment of the management platform. The product has passed CCC certification.
Frequently Asked Questions
Q1: How is the EC700's 320 TOPS heterogeneous computing power calculated? How many cards need to be added?
The EC700's standard RK3588J built‑in NPU provides 6 TOPS of compute power. The chassis has 2× M.2 PCIe high‑speed interfaces reserved for adding compute expansion modules. A single LQ50 module provides 160 TOPS@INT8, and dual modules stacked deliver 320 TOPS. The R182X module provides 20 TOPS@INT8 per unit, suitable for 3B to 7B parameter models; the LQ50 module provides 160 TOPS@INT8 per unit, suitable for 7B to 35B parameter models. Users can choose to install 1 or 2 modules as needed, achieving elastic compute upgrades.
Q2: How many large models can the EC700 run simultaneously? What is the inference latency for a 35B‑parameter model?
According to IOTRouter's internal performance testing data, under the dual LQ50 module configuration, the EC700 can simultaneously run inference workloads with two 35B‑parameter large models. Taking the Qwen3.5 35B‑A3B model as an example, under conditions of Input: 128 tokens and Output: 128 tokens, the TTFT (time to first token) is approximately 738 ms, TPOT (time per output token) is approximately 24.06 ms, and Decode TPS is approximately 24.06. The end‑to‑end latency is within an acceptable range, meeting the real‑time inference requirements of industrial edge AI.
Q3: What is the secondary development environment for the EC700? Which AI frameworks does it support?
The EC700 is based on an open embedded Linux system and provides a full development kit and API documentation. For AI inference, it supports mainstream deep learning frameworks such as TensorFlow, PyTorch, Caffe, and MXNet, as well as the ONNX universal model format, with flexible deployment based on Docker containers. The compute modules integrate a dedicated NPU plus the RK3588 ISP vision processing unit, natively supporting integrated multimodal inference across text, image, and video.
Q4: What industrial protocols does the EC700 support? Can it simultaneously connect to Siemens S7 and Mitsubishi FX?
The device supports over 100 industrial protocols via NeuronEX‑Lite, including Modbus RTU/TCP, OPC UA, EtherNet/IP, Siemens S7, Omron, Allen‑Bradley, and more. The device provides 2× RS485 + 1× RS232 serial ports, each configurable with different baud rates and protocols. This means a single EC700 can simultaneously communicate with multiple PLCs from different brands over different interfaces and protocols, with data unified and processed internally before being sent to the cloud.
Q5: Does the EC700's remote O&M require a public IP address? How is security guaranteed?
No public IP address is required. The EC700 has built‑in NAT traversal and P2P hole‑punching technologies and actively maintains a persistent connection to the cloud handshake server. When the O&M side initiates access, the cloud assists in establishing a point‑to‑point encrypted tunnel, reducing unnecessary cloud forwarding and improving latency and bandwidth efficiency. Together with IOTRouter's self‑developed free remote O&M software, it provides comprehensive capabilities for remote configuration, debugging, diagnostics, and updates — debugging PLCs deployed at remote sites feels much like being connected locally.
Learn More: Edge AI for Industrial Applications
Disclaimer: This article is based on publicly available product specifications and performance data from the manufacturer. All performance claims are sourced from the manufacturer's published test results.
Top comments (0)