Overview
Features
Qualcomm High-performance AI Processor IQ-9075
Integrated with an octa-core full-performance Kryo Gen6 CPU, with a maximum main frequency of 2.36 GHz and a general computing power of 230 KDMIPS. It has a built-in MCU like security subsystem equipped with four real-time cores.
200 TOPS Flagship-grade NPU
Dual Hexagon NPU, delivering up to 200 TOPS (sparse INT8) / 100 TOPS (dense INT8). It supports hardware collaborative acceleration of CPU/GPU/NPU, enabling private deployment of edge-side large models and generative AI, with large model inference performance up to 12 Tokens/s@13B.
Powerful GPU and 8K Encode/Decode VPU
Adreno 663 GPU, compatible with graphics and computing interfaces including Vulkan 1.2 and OpenGL ES 3.2. The VPU supports concurrent encoding and decoding: 1×8K60, 4×4K60, 32×1080P30, as well as simultaneous encoding and decoding of 2×4K60 + 2×4K60.
36GB LPDDR5 + 128GB UFS2.2
Integrated with 36GB LPDDR5 (ECC) high-speed memory and 128GB UFS2.2 flash storage, it also supports PCIe 4.0×4 M.2 NVMe SSD expansion, fully ensuring the high-speed reading and writing as well as stable operation of edge-side large model loading, high-definition videos and massive data.
Full-scenario Network Connectivity
Equipped with industrial-grade dual 2.5G Ethernet and TSN, the device supports optional expansion of Wi-Fi 6, Bluetooth 5.2, 4G and 5G. With low latency and high reliability, it perfectly meets the application requirements of industrial automation, real-time control and high-speed data backhaul.
Industrial-grade Design
Supports wide-temperature operation from -40℃ to 85℃, equipped with 8×GMSL2, 2×CAN-FD, 2×RS485, 12×optically isolated DI/DO and other interfaces, meeting expansion requirements for industrial control, robotics and other scenarios.
Complete Software Ecosystem
Relying on the Qualcomm Linux software stack, it natively supports Ubuntu and Yocto systems, is compatible with deep learning frameworks such as TensorFlow, PyTorch and ONNX, and can connect to model marketplaces and massive model repositories to quickly complete model conversion, tuning and validation.
Comprehensive Development Materials
Supporting source code, practical tutorials, detailed technical documents and development tools are provided to facilitate secondary development, streamline procedures and lower project development barriers.
Application
Edge Computing
Decentralizes computing to the edge for real-time local AI inference, minimizing latency and bandwidth costs for IoT endpoints.
Computing Power
Provides on-demand, scalable high-performance computing resources to accelerate AI training and data processing, optimizing IT infrastructure costs.
Industrial Automation
Integrates smart control and sensing for automated production, driving cost-effective digital transformation in manufacturing.
Smart Security
AI-powered visual analysis enables 24/7 monitoring and active anomaly alerts, shifting security from passive to proactive.
Specifications
| AIBOX-9075 KIT | ||
|---|---|---|
| Basic Specifications | SOC | Qualcomm IQ-9075 |
| CPU | Qualcomm Kryo™ Gen6 CPU: Octa-core 64-bit (8xKryo 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 2x Image Front End (IFE) + 5x 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 enhanced image processing features including target scaler, exclusion rectangle extraction, inline rotation, 17x17x17 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, H.265, VP9, MPEG2 Video encoding: 2×4K@60fps, 4×4K@30fps, 8×1080p@60fps, 16×1080p@30fps H.264, H.265, HEIF/HEIC | |
| RAM | 36GB LPDDR5 (ECC supported) | |
| Storage | 128GB UFS2.2 | |
| Storage expansion | 1 × M.2 (PCIe 4.0 x4 NVMe 2280 SSD expandable) | |
| 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 | 91.40mm × 102.10mm × 71.93mm (with heat sink) | |
| 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 × HD2.0 (4K@60Hz) | |
| USB | 2 × USB3.0 (Max: 1A), 1 × Type-C (Debug) | |
| Button | 1 × Power Button, 1 × Flash Button | |
| Others | 1 × SIM Card, 1 × FAN (2Pin-1.25mm), 1 × Phoenix terminal block (36Pin, 3.5mm pitch): 2 × RS485, 2 × CAN-FD, 12 × Opto-isolated DI/DO | |
