JetKU-CXP1004 Embedded Jetson CoaXPress AI Computing Platform
● Channel Specs: 4 independent CXP-12 channels, 12.5Gbps per lane, total bandwidth up to 50Gbps
● 40GQSFP+ optical port, 26-pin multi-function I/O
● Edge AI Performance: Up to 157 TOPS INT8 offline local inference, no external
● Applications: Industrial AOI inspection, scientific high-speed imaging, automation control, semiconductor test, mobile robot, medical & UAV onboard vision
JetKU-CXP1004 Embedded Jetson CoaXPress AI Computing Platform
Xilinx UltraScale+ FPGA + NVIDIA Jetson Heterogeneous Architecture | 4-Channel CXP-12 Edge AI Computing All-in-One | Integrated PoCXP Power Supply & Local Inference
Document Download
JetKU-CXP1004 4-Channel CoaXPress Frame Grabber User Manual.pdfFour Core Advantages
FPGA+Jetson Dual-Core Heterogeneous Computing Architecture
Equipped with Xilinx Kintex UltraScale+ XCKU5P FPGA + NVIDIA Jetson Orin Nano/NX SoM, interconnected via PCIe Gen3 x4 high-speed interface. The underlying FPGA handles image acquisition and preprocessing, while Jetson GPU performs local AI inference, enabling "acquisition and analysis in one" without the need for an industrial computer.
4-Channel CXP-12 Full Bandwidth Acquisition + PoCXP Power Supply
Supports 4 independent CoaXPress v2.1 camera inputs with a maximum speed of 12.5Gbps per channel (total aggregate bandwidth 50Gbps). Built-in PoCXP circuit delivers 17W 24V power per channel, transmitting data and power simultaneously over a single coaxial cable to simplify device wiring.
Complete Embedded Edge AI Software Stack
Pre-installed with Ubuntu 22.04 + NVIDIA JetPack SDK, including CUDA, cuDNN, TensorRT, OpenCV, and ROS support. Natively supports PyTorch/ONNX inference, paired with Aravis/GenTL standard SDKs, and offers dual control modes: local ImgGrab graphical debugging and remote upper computer control.
High Energy Efficiency Edge Computing Capability
Powered by NVIDIA Ampere architecture GPU, covering 34~275 TOPS INT8 inference computing power with a wide power consumption range of 7W~75W. Available in 4GB/8GB/16GB memory versions with modular SoM + carrier board design, easy to integrate into different products for various edge lightweight scenarios.
Product Overview
JetKU-CXP1004 is an embedded 4-channel CoaXPress (CXP) AI computing platform designed for edge imaging AI applications in industrial and scientific research fields. Compliant with CoaXPress V2.1 standard, a single Micro-BNC coaxial cable can synchronously transmit high-speed images and PoCXP camera power, significantly reducing the wiring complexity of the entire system.
Onboard Xilinx Kintex UltraScale+ XCKU5P FPGA serves as the underlying image processing unit, paired with NVIDIA Jetson Orin series GPU modules. The FPGA side features 2GB DDR4 SDRAM cache for real-time ISP, ROI cropping, and lossless compression. The Jetson side supports running deep learning models (e.g., object detection, semantic segmentation, speech recognition) locally without network dependency. Jetson Orin Nano is available in two versions: 4GB (34TOPS INT8, 7-25W) and 8GB (67TOPS INT8, 7-25W); Jetson Orin NX is available in two versions: 8GB (117TOPS INT8, 10-40W) and 16GB (157TOPS INT8, 10-40W). All models adopt NVIDIA Ampere architecture GPU, enabling all image algorithms, defect detection, and visual SLAM to run offline locally without an external industrial computer, highly suitable for lightweight edge scenarios such as mobile robots, medical devices, semiconductor AOI, and UAV airborne vision.
Core Features
- Core Architecture: Xilinx Kintex UltraScale+ XCKU5P FPGA + NVIDIA Jetson Orin Nano/NX AI SoM dual-processor heterogeneous computing architecture
- Acquisition Channels: 4 independent CoaXPress v2.1 inputs (backward compatible with CXP v2.0/v1.1), maximum 12.5Gbps per channel (CXP-12), total bandwidth 50Gbps
- PoCXP Power Supply: 17W 24V regulated power supply per CXP interface with overcurrent/short-circuit automatic protection, no external camera power supply required
- Internal Interconnection: PCIe Gen3 x4 connecting FPGA and Jetson, theoretical bandwidth 32Gbps for low-latency large data transmission
- Onboard Cache: 2GB DDR4 SDRAM on FPGA side; 4GB/8GB/16GB LPDDR4x on Jetson side (varies by model)
- AI Computing Power (Ampere Architecture GPU): Orin Nano 4GB (34TOPS,7-25W) / Orin Nano 8GB (67TOPS,7-25W) / Orin NX 8GB (117TOPS,10-40W) / Orin NX 16GB (157TOPS,10-40W)
- Edge AI Capability: Supports local execution of deep learning models without network dependency
- Local Storage: Standard M.2 NVMe 256GB SSD, supporting large-capacity image recording and local AI model deployment
- Industrial I/O: 26pin isolated industrial interface including RS485, differential LVDS, optocoupler isolated input/output, TTL encoder interface
- Expansion Interfaces: 40Gbps QSFP+ optical port, Gigabit Ethernet, USB3.2, Type-C, DP display output, Mini PCIe wireless slot
- System Environment: Pre-installed Ubuntu 22.04 LTS with complete NVIDIA JetPack AI software stack, supporting TensorRT accelerated inference
- Mechanical Specifications: Embedded computing platform with dimensions 186.00mm × 173.30mm, powered by 12V DC single power supply
Complete Technical Specifications
| Category | Item | Detailed Parameters |
|---|---|---|
| Core Computing Platform | FPGA Unit | Xilinx Kintex UltraScale+ XCKU5P, supporting user-defined logic development; 2GB DDR4 SDRAM large-capacity image cache |
| AI Processing Unit | • Jetson Orin Nano 4GB: 34TOPS INT8, 7-25W, 512-core Ampere GPU, max frequency 1020MHz (Entry-level robots/education/small AI devices) • Jetson Orin Nano 8GB: 67TOPS INT8, 7-25W, 1024-core Ampere GPU, max frequency 1020MHz (Entry-level robots/education/small AI devices) • Jetson Orin NX 8GB: 117TOPS INT8, 10-40W, 1024-core Ampere GPU, max frequency 1173MHz (Mid-range robots/Autonomous Mobile Robots (AMR)) • Jetson Orin NX 16GB: 157TOPS INT8, 10-40W, 1024-core Ampere GPU, max frequency 1173MHz (Mid-range robots/Autonomous Mobile Robots (AMR)) |
|
| CoaXPress Camera Input | Interface Standard | CoaXPress v2.1 (backward compatible with v2.0 / v1.0 / v1.1); 4 Micro-BNC channels CH0~CH3 |
| Downlink Speed | Full range of CXP-1(1.25G)~CXP-12(12.5G), maximum 12.5Gbps per channel, total aggregate 50Gbps | |
| PoCXP Power Supply | Maximum 17W 24V DC per channel with overload/short-circuit protection, onboard 12V to 24V DC-DC converter | |
| High-Speed Expansion Interfaces | QSFP+ Optical Port | 40Gbps high-speed optical fiber expansion, supporting multi-platform cascading and mass data offline export |
| Gigabit Ethernet | Remote SSH debugging, network video stream transmission | |
| USB3.2 Gen1 ×4 | Peripheral, dongle, high-speed storage device connection | |
| DP Display Output | Local HD interface preview and debugging display | |
| M.2 NVMe SSD | 256GB standard configuration for local image storage and AI model deployment | |
| Mini PCIe Slot | Reserved for expanding 4G/5G/LTE/Wi-Fi wireless modules (not included in standard configuration) | |
| General Industrial I/O (26pin DB26) | Differential Bus | 4 bidirectional LINE0-3 (RS485/422), 2 LVDS differential input/output |
| Isolated Circuits | 2 optocoupler isolated inputs, 2 optocoupler isolated outputs, withstand voltage 30V | |
| TTL GPIO | 2 bidirectional 3.3V LVTTL, supporting encoder A/B/Z signal acquisition | |
| Trigger Function | 2 32-bit hardware timers, event trigger, position trigger, noise filtering | |
| Software & Development Ecosystem | Operating System | Pre-installed on platform: Ubuntu 22.04 LTS; Upper computer: Windows7/10/11 64bit |
| AI Development Stack | NVIDIA JetPack SDK (including CUDA Toolkit, cuDNN, TensorRT, OpenCV, GStreamer), supporting PyTorch/ONNX Runtime and ROS | |
| Development SDK | Aravis SDK, GenTL Producer, ImgGrab graphical client, complete Qt demonstration project | |
| Power & Mechanical | Power Input | 12V DC input, 5A adapter recommended for powering the whole machine + PoCXP cameras |
| Dimensions & Heat Dissipation | 186mm × 173.3mm with active fan cooling; industrial-grade wide temperature operation |
NVIDIA Jetson Product Series Complete Parameter Reference (as of 2026)
| Module Model | AI Performance (TOPS) | Power Consumption Range | GPU Specification | GPU Max Frequency | Application Scenarios |
|---|---|---|---|---|---|
| Jetson AGX Orin Developer Kit | 275 | 15W–60W | 2048-core Ampere GPU, 64 Tensor Cores | 1.3 GHz | High-performance robots, autonomous driving, medical imaging |
| Jetson AGX Orin Industrial Edition | 248 | 15W–75W | 2048-core Ampere GPU, 64 Tensor Cores | 1.2 GHz | |
| Jetson AGX Orin 32GB | 241 | 15W–60W | 1792-core Ampere GPU, 56 Tensor Cores | 1.3 GHz | |
| Jetson Orin NX 16GB | 157 | 10W–40W | 1024-core Ampere GPU, 32 Tensor Cores | 1173 MHz | Mid-range robots, Autonomous Mobile Robots (AMR) |
| Jetson Orin NX 8GB | 117 | 10W–40W | 1024-core Ampere GPU, 32 Tensor Cores | 1173 MHz | |
| Jetson Orin Nano Super Developer Kit | 67 | 7W–25W | 1024-core Ampere GPU, 32 Tensor Cores | 1020 MHz | Entry-level robots, education, small AI devices |
| Jetson Orin Nano 8GB | 67 | 7W–25W | 1024-core Ampere GPU, 32 Tensor Cores | 1020 MHz | |
| Jetson Orin Nano 4GB | 34 | 7W–25W | 512-core Ampere GPU, 16 Tensor Cores | 1020 MHz |
Hardware Interface & Structure Drawings
Figure: Front Hardware Interface Labeled Diagram
Figure: Back Expansion Slot Interface Diagram
Figure: FPGA & Jetson Hardware Interconnection Architecture Diagram
Figure: Overall PCB Dimension Diagram
Software Architecture & Quick Start
Three Working Modes
- Edge Local Computing Mode (Recommended): Jetson independently completes image acquisition, AI inference, and motion control without external PC; deploy algorithm scripts via SSH terminal and debug graphically with local ImgGrab.
- Upper Computer Remote Control Mode: LAN PC connects to the computing platform remotely via GenTL/Aravis protocol for real-time image preview and camera parameter configuration, suitable for laboratory debugging.
- Local GUI Mode: Connect to DP display directly and run ImgGrab on the platform's Linux desktop to debug the entire visual system independently without a computer.
Complete Software List
| Category | Software Name | Usage Description |
|---|---|---|
| Underlying Driver | HelloFPGA Runtime Driver Package | Underlying hardware driver for computing platform, required for Windows upper computer |
| Upper Computer GUI | ImgGrab Client | Camera discovery, parameter configuration, image preview, video playback |
| Development SDK | GenTL SDK | Standard GenTL Producer cross-platform dynamic library |
| Aravis SDK | Complete Aravis 0.10 development library adapted to computing platform with Qt demonstration project | |
| Assistant Tool | Device Manager | Device information reading, online firmware upgrade |
All software can be downloaded from the official website: https://img-grab.com/jszc
- Hardware Wiring: Connect 12V 5A power adapter, CXP camera coaxial cable, and DP display/Gigabit network cable for debugging if needed;
- Power On: The computing platform is pre-installed with Ubuntu system, which automatically initializes Jetson and FPGA on first boot;
- Debugging Mode Selection: Open ImgGrab on local display directly, or install driver + ImgGrab on same-network PC for remote connection;
- AI Algorithm Deployment: Upload models via SSH, complete quantization acceleration with TensorRT, and run visual detection process locally in real time.
Safety Precautions (Must Read)
- Do NOT hot-plug CXP coaxial cables, FMC subcards, or power interfaces; operate hardware only after power off and electrostatic discharge;
- Maximum PoCXP output per channel is 17W; total power consumption of four cameras shall not exceed the rated power of the adapter, otherwise overcurrent protection will trigger power-off;
- Wear anti-static wristband before operation; the high-density FPGA and Jetson modules are extremely sensitive to ESD, which may cause permanent hardware damage;
- 26pin I/O interface level is 3.3V LVTTL; DO NOT connect 5V/12V external signals directly, otherwise FPGA IO circuits will be burned;
- Surface temperature of the whole machine can reach over 60℃ under full-load AI inference + 4-channel CXP-12 acquisition; touch and maintain only after power off and cooling;
- Store in anti-static bag in dry and ventilated environment when idle for a long time; power on regularly to remove moisture and avoid chip damage from condensation.
Typical Edge AI Application Scenarios
AMR/AGV Mobile Robot Visual SLAM Navigation
Semiconductor Wafer Defect Online AOI Inspection
UAV Airborne Electro-Optical Pod Visual Recognition
Portable Ultrasound/Intraoperative CT Medical Edge Imaging Computing
High-Speed Object Appearance Real-Time Sorting on Flexible Production Lines
High-Speed Vision Offline Acquisition Platform for Universities/Research Labs
