T-Chip Group
Intelligent hardware website
Innovate intelligent hardware and solutions.
Server Business Website
Server products and efficient solutions.
Consumer Business Website
StationPC NAS products and services.

AIBOX-9075 Industrial Edge Box

200 TOPS Computing Power NPU

· High-Performance AI Processor IQ-9075

· GPU/VPU Graphics & Multimedia Processing

· Up to 36GB LPDDR5 Memory

· All-Scenario Multi-Standard Network Connectivity 

· 8-Channel GMSL2 Vision Capture Interface

· Support -40°C to 85°C Wide Temperature

· Rich Industrial Expansion Interfaces

· Full-Stack Software Ecosystem Support


Manual Wiki Download Community Docs

AI Large Model

Ubuntu

Yocto

RTOS

AIBOX-9075
AIBOX-9075 Industrial Edge Box

Powered by Qualcomm IQ-9075 processor, delivering up to 200 TOPS of peak AI computing power and achieving 12TPS large model inference, enabling on-device private deployment of large language models. Its octa-core CPU runs at up to 2.36GHz, well-suited for high‑load edge AI and multi‑task concurrency. It features built‑in security subsystem and four real‑time cores. Leveraging Qualcomm Linux software stack, it supports Ubuntu/Yocto systems and mainstream AI inference frameworks. With an industrial wide‑temperature design, it comes standard with GMSL2, CAN‑FD, RS485, and opto‑isolated I/O.

AI Processor IQ-9075

Integrates octa-core big-core-only Kryo Gen6 CPU with a maximum frequency of 2.36 GHz, delivering 230 KDMIPS computing power. CPU/GPU/NPU co-acceleration is designed for high-load edge AI and multi-task. Features built-in MCU-like security subsystem with 4 real-time cores that can operate independently of application processor, enabling system monitoring, error detection, self-test diagnostics in real time.

Powerful GPU & 8K Codec VPU

Built-in Adreno 663 GPU supports 4K@60fps output and is compatible with mainstream graphics APIs including OpenGL ES, Vulkan, OpenCL, and OpenVX. Equipped with high-performance VPU video processing unit, it supports multi-channel concurrent encoding and decoding at 8K@60fps, 4×4K@60fps, and 32×1080P@30fps, enabling simultaneous encode and decode of 2×4K@60fps + 2×4K@60fps.

Up to 36GB LPDDR5 Memory

Onboard 36GB LPDDR5 (ECC) high-speed memory and 128GB UFS2.2 flash storage, with support for PCIe 4.0×4 M.2 NVMe SSD expansion. This fully ensures high-speed read/write performance and stable operation for on-device large model loading, high-definition video, and massive data processing.

Multi-Standard Network Connectivity

Equipped with dual 2.5G Ethernet RJ45 ports with support for TSN industrial real‑time networking. Wi-Fi 6 and Bluetooth 5.2 can be expanded via an M.2 slot, and 4G/5G wireless connectivity can be added via a Mini PCIe interface. With low‑latency and high‑reliability transmission characteristics.

8 x GMSL2 HD Vision Inputs

Equipped with 8 x GMSL2 high-speed vision interfaces, supporting simultaneous connection of HD cameras with a maximum transmission distance of up to 20 meters. Featuring strong anti-interference capability, it operates reliably in complex industrial electromagnetic environments, ensuring millisecond-level synchronization of multi-channel images.

-40°C~85°C Wide Temperature

Features unibody all-aluminum alloy enclosure with active cooling fan for efficient heat dissipation. It supports stable operation across ultra-wide temperature range of -40°C to 85°C, enabling 24/7 continuous operation. Built on rigorous industrial-grade reliability standards, it is perfectly suited for extreme high- and low-temperature environments such as in-vehicle installations, outdoor, and industrial facilities.

Industrial Expansion Interfaces

Equipped with 2 x CAN-FD, 2 x RS485, and 12 x opto-isolated DI/DO channels, fully adaptable to industrial field scenarios requiring highly reliable link connections, real-time data exchange, and multi-device coordinated control.

Full-Stack Software Ecosystem

Built on Qualcomm Linux — natively supports Ubuntu/Yocto, compatible with TensorFlow, PyTorch, ONNX, and more. Enables on-device private deployment of LLMs (DeepSeek-R1, Gemma, Llama, Qwen, Phi) and image generation models (Stable Diffusion). Connects to model hubs for fast conversion, tuning, and validation.
200 TOPS Flagship NPU Computing Power

Powered by Qualcomm IQ-9075 processor, featuring dual Hexagon NPUs with peak sparse INT8 computing power of up to 200 TOPS. It supports CPU/GPU/NPU hardware‑accelerated collaboration, enabling on‑device private deployment of large language models and generative AI, delivering LLM inference performance of up to 12TPS@13B. It efficiently handles a wide range of complex AI computations and large model inference tasks.

Application Scenarios

Widely applicable in industries such as robotics, private deployment of large models, edge computing, intelligent surveillance, industrial automation and smart cities.

Robots
Robots
Private Deployment of Large Models
Private Deployment of Large Models
Edge Computing
Edge Computing
Intelligent Surveillance
Intelligent Surveillance
Industrial Automation
Industrial Automation
Smart City
Smart City
Interfaces
Specifications

AIBOX-9075

Basic Specifications

SOC

Qualcomm IQ-9075

CPU

Qualcomm Kryo® Gen6 CPU: Octa-core 64-bit (8×Kryo Gold Prime), up to 2.36 GHz

MCU

Integrated subsystem with quad-core Cortex-R52 CPU; each Cortex-R52 CPU runs at up to 1.85 GHz, supporting independent boot via the OSPI interface

GPU

Adreno 663 GPU supports safe GPGPU compute, up to 840MHz · Graphics APIs: Vulkan 1.2, OpenGL ES 3.2 · Compute APIs: Vulkan 1.2, OpenCL 2.0 FP, Adreno NN Direct

ISP

Qualcomm Spectra 690 ISP Image Signal Processor, featuring 2× Image Front End (IFE) + 5× Lightweight Image Front End (IFE_L). Supports 24-bit HDR Bayer processing, lens distortion correction, advanced tone mapping, offset correction, lens vignetting correction, dead pixel correction, directional scaling, color lookup table, color space conversion and noise reduction.

NPU

Dual Hexagon Tensor Processor integrating Qualcomm Hexagon DSP, clocked up to 1.42 GHz. It incorporates 4 Hexagon Vector Extension (HVX) and 2 Hexagon Matrix Extension (HMX) cores, delivering up to 100 TOPS (Dense INT8) / 200 TOPS (Sparse INT8), and achieves 22 token/s for Llama2-7B inference.

DPU

Dual Adreno DPU1199 supports image processing features including target scaler, exclusion rectangle extraction, inline rotation, 17×17×17 3D LUT (ViG/DSPP), HDR10 enhancement, wide gamut WCG, corner rounding, and CCCS fixed-point conversion. It also supports UBWC 4.0 and DSC v1.2 image compression technologies.

Video codec

Video decoding: 1×8K@60fps, 2×8K@30fps, 4×4K@60fps, 8×4K@30fps, 16×1080p@60fps, 32×1080p@30fps AV1, H.264, VP9, MPEG2 Video encoding: 2×4K@60fps, 4×4K@30fps, 8×1080p@60fps, 16×1080p@30fps H.264, HEIF/HEIC

RAM

36GB LPDDR5 (ECC supported)

Storage

128GB UFS2.2

Storage expansion

1 × M.2 (PCIe 4.0 x4 NVMe 2280 SSD expandable; located inside the computer)

Power supply

DC 12V/5A (5.5 × 2.1mm)

Power consumption

Normal: 12W(12V/1000mA), Max: 60W(12V/5000mA), Min(Sleep): 1.08W(12V/90mA)

OS/Software

Ubuntu, Yocto Linux Software Stack

Software support

Supports private deployment of ultra-large parameter models under the Transformer architecture, including the Deepseek-R1 series, Gemma series, Llama series, Qwen series, Phi series and other large language models. Supports private deployment of image generation models such as Stable Diffusion. Supports the QNN AI inference framework, as well as multiple deep learning frameworks including TensorFlow, TensorFlow Lite, PyTorch, ONNX, Keras and Caffe.

Size

110.2mm × 102.25mm × 72.0mm (without mounting ears)

Environment

Operating Temperature: -40℃ ~ 85℃, Storage Temperature: -40℃ ~ 90℃, Storage Humidity: 10% ~ 90%RH (non-condensing)

Interface Specifications

Ethernet

2 × 2.5G Ethernet (2.5Gbps/RJ45)

Wireless network

Wi-Fi & Bluetooth: Supports Wi-Fi 6 and Bluetooth 5.2 (expandable via internal M.2 E-KEY) 4G/5G: Expandable via internal Mini PCIe

Video input

8 × GMSL2 (2×4Pin Mini FAKRA)

Video output

1 × HDMI2.0 (4K@60Hz)

USB

2 × USB3.0 (Max: 1A), 1 × Type-C (Debug)

Antenna

4 × 5G Antenna, 2 × Wi-Fi Antenna

Button

1 × Power Button, 1 × Burn Button

Others

1 × SIM Card, 1 × Phoenix terminal block (36Pin, 3.5mm pitch): 2 × RS485, 2 × CAN-FD, 12 × Opto-isolated DI/DO
Accessories

EC20 4G-LTE

$ 39.00

WXT14R2201 WiFi Module

$ 15.00

FORESEE M.2 2280 SSD

$ 79.00

Go to more product
Customization

Firefly team, with over 20 years of experience in product design, research and development, and production, provides you with services such as hardware, software, complete machine customization, and OEM server.

Hardware Solution

Software Solution

Turnkey Solution