In industrial and edge AI deployments, reducing Total Cost of Ownership (TCO) and optimizing Size, Weight, Power and Cost (SWaP-C) has always been a core challenge. Traditional architectures based on "x86 industrial PC + NVIDIA discrete GPU" often suffer from bulky size, excessive power consumption, and data privacy risks caused by cloud dependency. The newly launched FCU3502 Edge AI Terminal delivers a highly integrated heterogeneous computing architecture, offering a superior localized solution to these challenges.

As a strategic partner of Rockchip, Forlinx Embedded has long been deeply engaged in Rockchip platforms. The newly released FCU3502 high-performance AI edge computing terminal is designed with a heterogeneous architecture based on the RK3588 core platform and multi-type AI computing cards, providing high-performance AI computing power for edge scenarios. It provides 26 to 46 TOPS of computing performance, includes 36 proprietary algorithm libraries from Forlinx Embedded, and supports multiple large language models.

Thanks to this dual-computing architecture, the FCU3502 comprehensively outperforms traditional solutions in power consumption, size, cost and operating temperature range

Dual Computing Synergy: 36-Channel Video Analytics and Local LLM Parallel Operation

Addressing industry challenges such as inadequate device computing power, limited camera channel support, reliance on cloud computing, and high latency, the FCU3502 dual-computing AI vision scheme offers an effective solution.

The device features a classic combination of the Forlinx Embedded FET3588-C SoM and a computing expansion card, with dual chips working together to enhance system stability and maximize AI computing performance. The main controller RK3588 is equipped with an octa-core CPU and 6 TOPS NPU, responsible for overall system scheduling and task coordination, while efficiently supporting multi-channel video encoding and decoding to ensure long-term stable operation of the device. The expandable RK1828 computing card adds an extra 20-40 TOPS of dedicated hard-core computing power. With the dedicated hardware acceleration engine, its comprehensive performance far exceeds conventional devices on the market.

Many similar solutions on the market can only support up to 16 cameras per unit, requiring two devices to cover all monitoring points. In contrast, one FCU3502 unit can access up to 36 cameras, delivering the equivalent performance of two conventional devices. It eliminates the need for an extra unit, along with redundant deployment, joint debugging and maintenance work — bringing excellent cost advantages for customers.

Item	FCU3502 Performance Advantages
Multi-channel Access Capacity:	Maximum support for 36 channels of 720P@20FPS H.265 IP cameras to access simultaneously
Video Processing Capability:	Synchronously completes the full workflow including video streaming, decoding, preprocessing, inference, post-processing and display
Total Output Performance:	Supports concurrent multi-channel video processing, with a total output of approximately 300 FPS, maintaining 8FPS inference efficiency per channel under 36-channel load
Real-time Response Capability:	The end-to-end video processing latency is ≤1s, meeting real-time analysis requirements
Deployment Mode:	The full workflow runs in a local closed loop, no external inference server required, and no reliance on cloud network.

36-channel Video Analysis Effect Demonstration:

Local Offline LLM Operation for Private Deployment

With the support of the RK1828 computing expansion card, the FCU3502 further unlocks the operation capability of offline large language models and multimodal models. For scenarios with restricted network access or requirements that core data must not leave the internal network, the device can complete the full-link image-text multimodal inference locally, completely getting rid of the dependence on cloud networks.

A comparison of the actual performance of various mainstream LLM/VLM models:

Benefiting from the sufficient end-side computing power reserve and memory bandwidth of hundreds of GB/s, the device can maintain a generation speed of over 40 tokens when running 7B LLM and VLM models, and simultaneously support edge RAG retrieval Q&A based on private knowledge bases. Core data such as equipment manuals, process specifications, and inspection records are all stored locally throughout the entire process, and business workflows can operate in a complete closed loop within an isolated network. For customers with high sensitivity to data sovereignty, including the industrial, energy, critical infrastructure sectors and large enterprises, the FCU3502 is an excellent choice that balances performance and security.

Demonstration of the practical implementation effect of the energy storage cabin knowledge base Q&A project:

This video is a demonstration sample, and the background supports custom Logo and name settings.

Self-developed Vision Algorithm Library: Out-of-the-box Use, No Need for Training from Scratch

Forlinx Embedded provides a self-developed vision algorithm library that covers common recognition requirements in security, industrial and traffic scenarios, including 36 types of algorithms such as fire and smoke recognition, intrusion detection, fall detection, license plate recognition, and garbage detection. Relevant algorithms have completed pre-training and tuning, and can be called directly, eliminating the processes of algorithm R&D, computing power adaptation and model optimization, which helps shorten the cycle of project prototype verification and launch.

Algorithm Library Details:

Category	Detection Function
Face Recognition	Alarm for faces not looking straight ahead, face occlusion detection, face blacklist detection, repeated face swiping detection, face whitelist detection
Staff Status	Personnel counting, regional intrusion detection, personnel gathering detection, fall detection, personnel detention detection, mobile phone usage detection, personnel loitering detection, smoking detection, on-duty sleeping detection, no safety helmet wearing detection, off-post detection, no reflective vest wearing detection
Vehicle‑status	License plate color recognition, electric vehicle illegal parking, passing vehicle counting, license plate recognition, fire channel illegal parking, license plate blacklist, vehicle intrusion, parking space occupation, non-local license plate recognition, non-motor vehicle intrusion, electric bicycle entering the room, engineering vehicle illegal entry, long-term motor vehicle detention, motor vehicle illegal parking, regional vehicle quantity over-limit detection
Object Detection	Garbage detection, fire extinguisher missing detection, smoke detection, open flame detection

Complete supporting services are available: self-developed algorithms are ready for immediate use, and operation guidelines for self-importing algorithms and support for customized algorithm projects are also provided. It saves the R&D cycle and cost of training from scratch. The complete solution adopts a full-stack architecture design, which is interconnected layer by layer from the hardware layer, algorithm layer, application layer to the system layer, realizing end-to-end full capability delivery.

On the camera side, it widely supports standard RTSP and ONVIF protocols, so existing IP cameras and mainstream platforms can be directly connected without additional adaptation. The management background provides a responsive Web GUI, which can complete algorithm configuration, camera management, real-time preview and alarm viewing in one stop.

Camera Management Demonstration Video:

Consistent Stability: Adaptable to On-site Deployment Environments

Edge computing terminals are usually deployed in on-site environments such as equipment rooms, distribution boxes and production lines, which put forward high requirements for the structural adaptation and heat dissipation capacity of the equipment.

FCU3502 Thermal Simulation Effect Display

The overall dimension of FCU3502 is 232.9mm×140.6mm×65.4mm, supporting two installation methods: desktop type and ear-mounted type. The complete machine has passed the temperature reliability test verification from 0℃ to 70℃, and the RK1828 can still maintain full-load operation under high-temperature working conditions, ensuring stable and continuous output of computing performance.

This all-in-one device features Forlinx Embedded's unified new design, topped with a programmable RGB light bar. Dedicated light effects intuitively signal operational and alarm status, letting maintenance staff spot equipment health from meters away — no close-range checks needed, for faster, easier inspections.

The FCU3502 features dual Gigabit Ethernet ports, USB 3.0/USB 2.0, multiple RS232/RS485 ports, digital input/output (DI/DO), HDMI input/output and 4G/5G antennas, enabling direct connection to on-site equipment.

Schematic diagram of on-site deployment

Function	Quantity	Description
Ethernet	2	Supports data transmission rates of 10/100/1000 Mbps;
USB2.0	2	Routed out via Type-A, 5V/0.5A output;
USB3.0	2	Routed out via Type-A, 5V/1A output;
HDMI TX	1	Connected via a standard HDMI socket; Supports up to 7680x4320@60Hz;
HDMI RX	1	Connected via a standard HDMI socket; Supports up to 4K@60Hz;
TF	1	Supports TF card to expand storage and flashing operating system;
DEBUG	1	RJ45 connector, RS232 level, connect to the computer through the console debugging line for debugging;
Audio	1	Built-in speakers and supports headphone output.
WiFi&BT	1	Supports WIFI4 & Bluetooth 4.2
M.2	1	4G/5G（EM05/RM500U）
	1	M.2 solid-state drive
	1	RK1828 computing power card.
RS485	2	Isolated RS485 level converter using a 3Pin 15EDGK-3.81 green terminal header.
RS232	4	Isolated RS232 level converter using a 3Pin 15EDGK-3.81 green terminal header.
DI	2	2 x optocoupler-isolated signal input, supporting dry contact connection.
DO	2	2 x relay-isolated output, supporting dry contact output.
KEY	4	4 functional keys: Reset, Power On/Off, MASKROM, and RECOVERY.
LED	2	1 always-on power supply status LED.

To meet on-site deployment reliability requirements, this product is being certified in compliance with CE, FCC and RoHS standards, and has completed relevant environmental tests, with certification completion scheduled for Q4.

For projects requiring concurrent multi-channel video analysis and local large model inference, the FCU3502 high-performance AI edge computing terminal delivers a cloud-independent, localized solution: decoding, streaming and inference are all executed locally on the device.

No data leaves the internal network, enabling end-to-end business operation in isolated network environments. The FCU3502 high-performance AI edge computing terminal is now officially launched. Full technical specifications and selection documentation are available via Forlinx Embedded's official channels.

---