The Changfan Tech CF-28F1E4 is a flagship AI computing platform engineered for next-generation high-compute edge applications. It supports the NVIDIA® Jetson AGX Thor series modules (T5000 & T4000), delivering a staggering AI performance of up to 2070 TFLOPS (FP4-sparse). Featuring the latest Blackwell architecture GPU, up to 128GB of LPDDR5X memory, and highly reliable industrial interfaces, the CF-28F1E4 effortlessly handles local Edge LLM inference, real-time analysis of multiple 8K ultra-HD video streams, and extremely complex deep learning tasks.
Engineered for Embodied AI and Harsh Environments
Designed for industrial sites and autonomous driving/robotics applications, the CF-28F1E4 features a wide-temperature design supporting -25°C to +65°C. It offers a wide DC 9V~36V power input, built-in 4x isolated GPI and 4x isolated GPO (3.3V/5V adjustable), and 3x isolated UARTs to ensure signal integrity and anti-interference capabilities in complex electromagnetic environments. The internal Nano SIM slot provides flexible 4G/5G wireless remote access capabilities.
Application Scenarios
Edge LLM Inference: Running large language models and vision models smoothly locally, thanks to the massive memory (up to 128GB) and ultra-high computing power.
Embodied AI & Advanced Robotics: Processing complex real-time motion control, environmental perception, and multi-sensor fusion (LiDAR, cameras, radar).
Advanced Machine Vision (AOI): Real-time decoding of multiple 8K/4K video streams and defect detection.
Autonomous Driving & V2X: Serving as a roadside edge computing node or a high-level autonomous driving test platform.
Smart Healthcare & Life Sciences: Gene sequencing analysis and real-time processing of high-end medical imaging.
🛠 Changfan Exclusive Customization Services
As a professional edge computing hardware manufacturer, Changfan Tech offers comprehensive customization support from proof-of-concept to mass production, precisely matching your project requirements:
Hardware & I/O Customization: Modifying specific I/O interfaces, adjusting internal expansion slots (e.g., adding miniPCIe or M.2 modules), or customizing thermal solutions based on project needs.
Mechanical & Enclosure Customization: Adjusting chassis dimensions, designing dedicated mounting brackets, and supporting custom brand logo printing and color finishing.
Software & OS Pre-configuration: Customizing Ubuntu OS images, pre-installing NVIDIA JetPack SDKs, Docker containers, or pre-flashing your proprietary AI models and algorithms.
Flexible Module Configuration: Offering Jetson AGX Thor T5000 and T4000 module options to precisely match your computing power, memory, and budget requirements.
Specifications Table
Module | Jetson Thor T5000 | Jetson Thor T4000 |
|---|---|---|
AI Performance | 2070 TFLOPS (FP4—sparse) | 1200 TFLOPS (FP4 — sparse) |
GPU | 2560-core NVIDIA Blackwell architecture GPU with 96 fifth-gen Tensor Cores | 1536-core NVIDIA Blackwell architecture GPU with 64 fifth-gen Tensor Cores |
CPU | 14-core Arm® Neoverse®-V3AE 64-bit CPU | 12-core Arm® Neoverse®-V3AE 64-bit CPU |
Memory | 128 GB 256-bit LPDDR5X | 64 GB 256-bit LPDDR5X |
Storage | Supports NVMe through PCIe | Supports NVMe through PCIe |
Video Encode | 6x 4Kp60 (H.265), 12x 4Kp30 (H.265), 24x 1080p60 (H.265) | 6x 4Kp60 (H.265), 12x 4Kp30 (H.265), 24x 1080p60 (H.265) |
Video Decode | 4x 8Kp30 (H.265), 10x 4Kp60 (H.265), 22x 4Kp30 (H.265) | 4x 8Kp30 (H.265), 10x 4Kp60 (H.265), 22x 4Kp30 (H.265) |
Display Interface | 1x HDMI 2.0 | 1x HDMI 2.0 |
USB Connector | 2x USB Type-C, 4x USB Type-A | 2x USB Type-C, 4x USB Type-A |
Ethernet | 4x RJ45 | 4x RJ45 |
Buttons | 1x Recovery, 1x Reset, 1x PWR | 1x Recovery, 1x Reset, 1x PWR |
Expansion Interface | 1x miniPCIe, 1x M.2 M key (2280) | 1x miniPCIe, 1x M.2 M key (2280) |
I/O Interface | 2x AUDIO, 2x MIC | 2x AUDIO |
Serial Port | 2x RS232, 2x RS485, 3x UART 3.3V | 2x RS232, 2x RS485, 3x UART 3.3V |
Operating Temperature | -25℃ ~ +65℃ | -25℃ ~ +65℃ |
Dimensions | 187mm x 180mm x 75mm | 187mm x 180mm x 75mm |
Power Requirement | DC + 9V~+36V | DC + 9V~+36V |
Weight | 1763g | 1763g |
Q1: How capable is the CF-28F1E4, powered by NVIDIA Jetson AGX Thor, in running Large Language Models (LLMs) at the edge?
A: The CF-28F1E4 is purpose-built for edge-based large model deployment. Equipped with the latest Blackwell architecture GPU, it delivers a staggering computing power of up to 2070 TFLOPS (FP4-sparse) and features up to 128GB of LPDDR5X memory (273 GB/s bandwidth). This combination of massive memory and ultra-high computing power allows it to smoothly run complex, tens-of-billions-parameter LLMs and Vision-Language Models (VLMs) completely offline in industrial environments, achieving localized AI inference with extremely low latency.
Q2: What are the generational improvements of this computing platform compared to previous-generation devices based on Jetson AGX Orin?
A: Compared to the Orin architecture (up to 275 TOPS), the Thor platform (T5000/T4000) on the Changfan CF-28F1E4 represents a quantum leap:
Computing Power Leap: Increased from 275 TOPS to up to 2070 TFLOPS (FP4).
Architecture Upgrade: Upgraded from Ampere to the latest Blackwell architecture with 5th-gen Tensor Cores.
Doubled Memory: Memory capacity increased to up to 128GB LPDDR5X, perfectly solving the "VRAM anxiety" for large model deployment.
Video Processing: Supports video decoding up to 4x 8Kp30, meeting the demanding requirements of multi-channel ultra-HD visual analysis.
Q3: What industrial-grade interfaces does the CF-28F1E4 feature, and how does it meet the requirements for Sensor Fusion?
A: To seamlessly connect with complex industrial environments, the CF-28F1E4 provides an extremely rich I/O configuration:
Networking & Cameras: 4x Gigabit RJ45 ports and 6x USB 3.0/3.2 ports ensure high-speed access for multiple HD industrial cameras.
Industrial Control: Integrates 4x CAN buses, 2x RS232, 2x RS485, and 3x isolated UARTs, capable of directly connecting to PLCs, servo drives, and LiDARs.
Isolation Protection: Built-in 4x electrically isolated GPI and 4x electrically isolated GPO (3.3V/5V adjustable) effectively prevent electromagnetic interference.
This makes it highly suitable for scenarios requiring the processing of multi-source data, such as Embodied AI robots, autonomous driving testing, and advanced machine vision.
Q4: What OEM/ODM customization services does Changfan Tech offer for the CF-28F1E4?
A: Changfan Tech provides comprehensive customization from Proof of Concept (POC) to mass production:
Hardware Modifications: Add or remove specific I/O interfaces, or adjust internal miniPCIe / M.2 expansion slots based on project requirements.
Structural & Appearance Customization: Custom chassis dimensions, dedicated mounting brackets, and support for customer brand logo printing and color customization.
Software Pre-installation: Supports factory pre-installation of customized Ubuntu OS images, NVIDIA JetPack SDKs, or even pre-flashing your proprietary AI algorithm models.
Q5: What are the wide-temperature and wide-voltage design specifications of the device? Does it support pre-installing the JetPack SDK?
A: The CF-28F1E4 features a rigorous industrial-grade design, supporting a wide operating temperature range of -25°C to +65°C and a wide voltage input of DC 9V~36V. On the software side, the device supports pre-installing a clean Ubuntu system and fully supports the installation of the NVIDIA JetPack SDK, allowing developers to directly utilize toolchains like CUDA and TensorRT for model acceleration.
Q6: As a B2B customer, does purchasing the CF-28F1E4 include technical support and warranty services?
A: Absolutely. As a professional edge computing hardware manufacturer, Changfan Tech provides all B2B customers with complete technical documentation (including interface definitions, dimension drawings, Recovery mode instructions, etc.) and SDK testing sample programs (such as CAN/RS485/GPIO test codes). The device comes with a standard original manufacturer hardware warranty, and we provide dedicated FAE (Field Application Engineer) technical support for bulk purchases and OEM projects.
