RK3576 + RK1828 Edge AI Development Kit for Local LLM, VLM & Vision (20 TOPS, 5GB DRAM, IMX415 Camera)

The RK3576 + RK1828 AI Development Kit combines an RK3576 host with a 20 TOPS RK1828 AI coprocessor featuring 5GB dedicated DRAM and up to 1TB/s memory bandwidth for edge AI development and evaluation.

The RK3576 and RK1828 hardware and software environment has been tested by DFRobot, reducing the work required for host selection and compatibility testing. The platform supports LLM, VLM, speech and computer vision workloads for robotics, industrial vision and intelligent edge devices.

20 TOPS AI Acceleration with 5GB Dedicated High-Bandwidth Memory

The RK1828 is a dedicated AI coprocessor designed for edge inference workloads. It provides 20 TOPS INT8 AI computing performance and integrates 5GB of dedicated DRAM with up to 1TB/s memory bandwidth, providing dedicated compute and memory resources for large language, vision-language and computer vision models.

The available RK1828 model ecosystem includes models such as Qwen3-8B, Qwen2.5-7B, Qwen2.5-VL-7B, InternVL, Gemma, Whisper, DINO, SigLIP and YOLO, enabling developers to evaluate workloads ranging from local conversational AI to visual understanding and real-time vision inference.

Skip Host Selection and Compatibility Testing

An AI accelerator is only useful when the host platform, PCIe interface, power system, drivers and software environment work together. With a standalone RK1828 module, developers need to verify whether their existing host provides a compatible M.2/PCIe interface, suitable power delivery and supported software environment.

This kit pairs the RK1828 with a validated RK3576 development platform, giving developers a known hardware configuration for evaluating the accelerator.

RK3576

• 8GB System Memory
• 64GB eMMC
• Debian Linux Environment
• M.2 PCIe Connectivity for RK1828
• Camera, Display, Network and Industrial Interfaces
• Tested RKNN-based Development Environment

This allows engineering teams to spend less time troubleshooting host compatibility and more time evaluating AI performance and developing applications.

Broad Model Support across LLM, VLM, Speech and Vision

The RK1828 software ecosystem covers multiple major edge AI workload categories, allowing developers to evaluate different model architectures using the same accelerator platform.

• Large Language Models: Qwen2.5 / Qwen3 / Gemma / GLM Edge / Tencent Youtu LLM
• Vision-Language Models: Qwen2.5-VL / Qwen3-VL / InternVL / FastVLM / Gemma Multimodal / SmolVLM / PaddleOCR-VL
• Speech AI: Whisper / Qwen ASR / Qwen TTS / SenseVoice / VITS / Zipformer
• Embedding & Retrieval: Qwen Embedding / Qwen Reranker / GME Qwen
• Computer Vision: YOLOv5 / YOLOv6 / YOLOv8 / MobileNet / ResNet50 / DINOv3 / SigLIP / Depth Anything V2

With pre-converted model resources and RKNN3 deployment workflows available for supported models, developers can begin evaluation without building every model conversion and quantization workflow from scratch.

Move from Hardware Setup to Working AI Examples Faster

The platform provides tested driver resources, model files and application examples designed to shorten the path between installing the hardware and seeing a working AI result.

DFRobot provides RKNN examples and software resources covering workloads such as YOLO object detection, instance segmentation, pose estimation, CLIP image-text matching, Whisper speech recognition, TTS, OCR and multimodal inference.

For RK1828 workloads, supported model and demo resources are provided through the product documentation. Once the required driver and model resources are prepared, provided examples can be launched directly instead of requiring developers to build an application from scratch.

An Integrated Edge AI Prototyping Platform

The RK3576 host platform provides the connectivity and interfaces required to turn AI inference into a working application prototype.

Kit Platform

• RK3576 Host with 8GB Memory and 64GB eMMC
• RK1828 20 TOPS AI Coprocessor
• IMX415 Camera Module
• 2 RGMII Ethernet Interfaces
• HDMI and DP Display Output
• MIPI-DSI Display Interface
• Multiple MIPI-CSI Camera Interfaces
• USB 3.2 and USB 2.0 Connectivity
• M.2 B-Key Expansion
• CAN FD / RS485 / UART
• Wi-Fi Connectivity

This allows developers to connect cameras, displays, communication modules, sensors and other peripherals when moving from model evaluation to application prototyping.

Start Vision AI Development with an Included Camera

An IMX415 camera module is included with the kit, providing developers with a hardware input for testing computer vision workloads.

Combined with the RK3576 camera interfaces and available YOLO, OCR, face detection and vision-language examples, the camera can be used to evaluate visual perception pipelines without selecting a separate camera module first.

Rockchip RK1828 Performance Reference

Rockchip Reference Benchmarks across LLM, VLM and Computer Vision Workloads. The following figures are Rockchip reference benchmark results under the specified test conditions. They are provided as accelerator-level performance references and are not measured performance results of this RK3576 + RK1828 development kit.

LLM Performance

Test Conditions: 128 input tokens and 128 generated tokens.

How to read the results: TTFT (Time to First Token) measures the delay before the first generated token appears. TPOT (Time per Output Token) measures the average generation time per output token. Decode TPS represents token generation throughput during the decoding stage.

VLM Performance

Test Conditions: The vision encoder and LLM are benchmarked separately. Vision input resolution is 384 × 384. LLM testing uses 128 input tokens and 128 generated tokens.

Note: Vision Latency and LLM Decode TPS measure different stages of the VLM pipeline and should not be interpreted as end-to-end multimodal application latency.

CNN Performance

Test Conditions: CNN benchmarks use 640 × 640 input resolution. Single-Core FPS represents single-core inference performance, while Multi-Batch Multi-Core FPS represents throughput using multiple NPU cores and batched workloads.

Note: Multi-Batch Multi-Core FPS is a throughput metric and should not be interpreted as the frame rate of a single-stream end-to-end vision application.

Benchmark Note: These tables are Rockchip reference benchmark results. The Rockchip reference platform uses RK3588 + RK1828 over PCIe unless otherwise specified in the source benchmark. Actual performance on this RK3576 + RK1828 kit may vary depending on the host platform, RKNN3 SDK/runtime version, model configuration, quantization, context length, batch size, input/output settings and application workload.

Features

• RK3576 host platform with quad-core Cortex-A72 @ 2.2GHz, quad-core Cortex-A53 @ 1.8GHz and Cortex-M0
• RK3576 integrated NPU delivers 6 TOPS @ INT8 computing power
• RK1828 AI coprocessor delivers 20 TOPS @ INT8 computing power
• RK1828 integrates 5 GB DRAM with 1 TB/s bandwidth for AI inference workloads
• 8 GB LPDDR5 and 64 GB eMMC for local development and model storage
• Supports LLM, VLM, ASR, TTS, retrieval and computer vision model workloads
• Supports mainstream model families including Qwen, Gemma, InternVL, Whisper, SigLIP, DINOv3, MobileNet, ResNet50 and YOLO
• RKNN3 toolchain supports model conversion, inference, performance evaluation and board-side deployment
• Pre-converted selected models, deployment examples and RK1828 demo resources available through the product Wiki
• Extensive expansion interfaces including HDMI, DP, MIPI-DSI, MIPI-CSI, USB, M.2, CAN FD, RS485 and UART

Applications

• Evaluate Edge AI Architectures Before Product Development: technology evaluation, algorithm verification and prototype development.
• Robotics: Evaluate local language, vision and multimodal AI for robot perception, interaction and task understanding.
• Industrial Vision: Prototype AI pipelines for object detection, defect inspection, OCR, image classification and visual analysis.
• Intelligent Edge Devices: Evaluate local LLM, VLM, speech and multimodal capabilities for smart terminals, meeting devices, AI appliances and edge computing systems.

Specifications

Basic Parameters

• System-on-Chip (SoC): RK3576 Module
  CPU: Quad-core Cortex-A72 @ 2.2GHz + Quad-core Cortex-A53 @ 1.8GHz + Cortex-M0
  Memory and Storage: 8GB LPDDR5 + 64GB eMMC
  NPU Compute Performance: 6 TOPS @ INT8
• AI Co-Processor: RK1828 (M.2) Module
  DRAM: 5GB
  DRAM Bandwidth: 1TB/s
  NPU Compute Performance: 20 TOPS @ INT8
• Power Supply:
  DC 2.5 Barrel Jack: 12V DC input
  Terminal Block: 12V DC input
• Required External Power Supply for AI Inference: 12V / 4A or higher (not included)
• Operating System: Debian
• Dimensions: 146 × 102 mm

Interface Specifications

• Video Output:
  1 × HDMI
  1 × DP
  1 × MIPI-DSI
• Video Input: 3 × MIPI-CSI
• M.2:
  1 × M.2 M Key (PCIe 2.0 x2)
• Communication Interfaces:
  2 × RGMII
  1 × Wi-Fi
  1 × M.2 B Key
• USB:
  1 × USB 3.2 OTG
  1 × USB 3.2 Host
  2 × USB 2.0 Host
• Audio:
  1 × Microphone input
  1 × 2W power amplifier
  1 × Headphone jack
• Other Interfaces:
  2 × CAN FD
  1 × RS485
  4 × UART

Package contents

• RK3576 Development Board x1
• RK3576 Heatsink (with mounting accessories) x1
• IMX415 Camera Module x1
• 2.4/5.8GHz FPC Antenna x1
• RK1828 AI Accelerator (M.2 Module) x1
• 15cm Single-head GH1.23-4P Cable x1

Important: A compatible 12V/4A or higher power supply is required and is not included. The RK1828 module requires simple installation into the RK3576 M.2 slot before use.