Rockchip RK1828 AI Acceleration Module for Edge AI (M.2, 20 TOPS)

This RK1828 AI accelerator is a high-performance AI coprocessor designed to add local AI inference capability to edge computing systems. Built on the Rockchip RK1828 platform, this M.2 AI accelerator integrates a 20 TOPS NPU with 5 GB of high-bandwidth DRAM, providing dedicated acceleration for large language models, multimodal models, speech processing and computer vision workloads in robotics, industrial systems, smart terminals and edge servers.

20 TOPS Compute with 1 TB/s Memory Bandwidth

This edge AI accelerator delivers 20 TOPS of NPU performance at INT8 and supports INT4, INT8, INT16, FP8, FP16 and BF16 precision formats. The integrated 5 GB DRAM provides 1 TB/s bandwidth, reducing memory-access bottlenecks during model inference. Under the specified test conditions, Qwen3-8B reaches 61.34 decode tokens per second, while Qwen3-4B reaches 88.47 tokens per second, enabling responsive local large-model inference for AI assistants and human-machine interaction.

One Accelerator for LLM, VLM, Speech and Vision Workloads

This RK1828 NPU supports more than 20 mainstream models across LLM, VLM, retrieval, ASR, TTS and computer vision categories. Supported examples include Qwen3-8B, Qwen2.5-VL-7B, Qwen2.5-Omni-3B, InternVL3.5-4B, Whisper, Qwen3-ASR, Qwen3-TTS, YOLOv8s and DINOv3. Pre-converted RKNN models and sample code reduce the amount of model conversion and deployment work required for proof-of-concept development.

RKNN3 Toolchain for Model Conversion, Deployment and Benchmarking

This AI coprocessor works with the RKNN3 software stack for model conversion, inference execution, performance evaluation and board-side deployment. Support for frameworks including TensorFlow, Caffe, TFLite, PyTorch, ONNX NN and Android NN provides a practical path from existing AI models to hardware-accelerated edge deployment.

Standard M.2 Form Factor for Edge System Integration

This M.2 AI accelerator uses a B-M Key interface and communicates through PCIe 2.1 ×1 Lane. The module follows the M.2 2280 form factor and uses an external DC 12 V power input, reducing the power demand placed on the host M.2 slot. Android, Linux and Windows compatibility allows integration with x86 or ARM hosts, industrial PCs and compatible development boards.

With dedicated AI compute, high-bandwidth memory and support for multiple AI model categories, this RK1828 AI accelerator is suitable for upgrading edge platforms that require local inference. Typical deployments include home computing hubs, industrial defect inspection, robot control systems, security surveillance and meeting transcription systems.

Features

• 20 TOPS NPU computing performance at INT8
• 5 GB integrated DRAM with 1 TB/s memory bandwidth
• Qwen3-8B decode performance of 61.34 tokens/s under specified test conditions
• Supports INT4 / INT8 / INT16 / FP8 / FP16 / BF16 precision formats
• Supports LLM, VLM, retrieval, ASR, TTS, image feature, classification and object detection models
• RKNN3 toolchain supports model conversion, inference, benchmarking and deployment
• M.2 B-M Key interface with PCIe 2.1 ×1 Lane communication
• Compatible with Android, Linux and Windows systems
• External DC 12 V power input reduces power demand on the host M.2 slot

Applications

• Home computing hubs
• Industrial defect inspection
• Robot control cores
• Security surveillance
• Meeting transcription systems

Specifications

Basic Parameters

• VCC_IN Input: DC 12 V
• Power Consumption:
  Average: 13.6 W
  Peak: 29.4 W (under 7B model workload)
• Hardware Interface: M.2 B-M Key
• Communication Bus: PCIe 2.1 ×1 Lane
• NPU Compute Performance: 20 TOPS @ INT8
• Supported Precision Formats: INT4 / INT8 / INT16 / FP8 / FP16 / BF16
• Supported Frameworks: TensorFlow, Caffe, TFLite, PyTorch, ONNX NN, Android NN, etc.
• DRAM Configuration: 5 GB, 1 TB/s bandwidth

Physical Dimensions

• Dimensions: 25 × 81 × 21.75 mm (heat sink included)

Package contents

• RK1828 AI Accelerator (M.2 Module) x1
• Single-head GH1.23-4P 15 cm Cable x1