A single board computer with built-in AI, PCIe support, and camera input allows local processing for vision, robotics, and edge devices without cloud use.
Radxa has launched the Cubie A7S, a 51 × 51 mm single board computer based on the Allwinner A733. It uses Cortex-A76 and Cortex-A55 CPU cores, LPDDR5 memory, PCIe 3.0 expansion, and a built-in NPU rated at 3 TOPS. The board is designed for edge AI, vision systems, and embedded multimedia use where small size, on-device processing, and low power use are needed.
The Cubie A7S helps solve a common issue in embedded design. Many developers need local AI inference without cloud connectivity. Sending data to the cloud can increase delay, bandwidth cost, and privacy risk. By combining CPU, GPU, NPU, and real-time control in a small board, it allows on-device processing for vision, analytics, and control systems. It is suitable for OEMs, system integrators, robotics developers, AI startups, and engineers building space-limited products.
The Allwinner A733 uses an octa-core setup with two Cortex-A76 cores running up to 2.0GHz and six Cortex-A55 cores up to 1.8GHz. Graphics are handled by an Imagination PowerVR BXM-4-64 MC1 GPU. It supports OpenGL ES 1.1, 2.0, and 3.x, Vulkan 1.3, and OpenCL 3.0 for graphics and compute tasks.
The NPU delivers up to 3 TOPS at INT8 precision and supports INT4, INT8, INT16, FP16, and BF16 data formats. The platform works with AI frameworks such as TensorFlow, PyTorch, ONNX, Caffe, TFLite, and Darknet. It supports Linux and Android, allowing model deployment across platforms and faster development.
Along with the Arm CPU cluster, the SoC includes a RISC-V XuanTie E902 microcontroller running up to 200MHz. This MCU manages low-power and real-time tasks such as motor control, sensor data capture, and system management when the main cores are in low-power mode. This setup allows control tasks to run separately from AI workloads.
The board supports LPDDR5 memory up to 16GB. Storage options include onboard eMMC up to 256GB and a microSD card slot. A PCIe 3.0 x1 interface is available through an FPC connector for NVMe SSD expansion, useful for data logging and AI model storage.
For video, the board supports H.264 and H.265 encoding up to 4K at 30fps. It also supports hardware decoding for H.265, VP9, and AVS2 up to 8K at 24fps. Display output is available through a USB-C port with DisplayPort Alt Mode, supporting up to 4K at 60fps. A 4-lane MIPI CSI interface allows direct camera connection for machine vision and surveillance systems.
