RK1820/RK1828 M.2 Computing Card

Product model: RK1820 / RK1828

Product interface: Standard M.2 M-Key

Product features: High-Performance NPU: 20 TOPS (INT8) with INT4/INT8/FP16 mixed-precision.

Brief Message Compare

Overview
Specification
Documentation
Order
Filter

RK1820 / RK1828 M.2 Computing Card

Empower your edge devices with the Rockchip RK1820/RK1828 M.2 Computing Card. Designed as a dedicated AI co-processor, it integrates a powerful 20 TOPS NPU and up to 5GB of 3D stacked in-package DRAM. This plug-and-play accelerator smoothly handles 3B-7B parameter LLMs and VLMs offline, offering unparalleled performance, high bandwidth, and privacy for localized edge AI workloads.

Highlights:

Robust AI Compute: 20 TOPS (INT8) NPU supporting mixed-precision ( INT4/INT8/FP16) for an optimal balance of accuracy and efficiency.
3D Stacked Memory: Built-in 2.5GB (RK1820) or 5GB (RK1828) high-bandwidth DRAM, eliminating external DDR bottlenecks and the edge AI "memory wall."
Offline GenAI Inference: Independently runs 3B-7B LLMs/VLMs (e.g., Qwen, LLaMA2) at 100+ tokens/s for low-latency operation without cloud dependency.
Decoupled Architecture: Standard M.2 2280 (M-Key) interface utilizing PCIe 2.1 1-Lane. Acts as a dedicated co-processor to prevent host CPU, memory, and bandwidth contention.
Broad Ecosystem: Plug-and-play compatibility with RK3588 , RK3576, RK3568, and RK3572 hosts. Natively supports RKNN, TensorFlow, PyTorch, and ONNX.

AI Accelerator M.2 Computing Card

https://www.forlinx.net/file.php?f=202608/f_f5d4580109e40f57403edf96385ea3c3&t=webp&o=&s=&v=1787037077 https://www.forlinx.net/file.php?f=202608/f_2c40ef1e6b75bb3d7e6e51e01f9dcf45&t=webp&o=&s=&v=1787037342 #F5F5F7 RK1820/RK1828 M.2 Computing Card Integrates a 20 TOPS (INT8) NPU and up to 5GB 3D stacked in-package DRAM. Enables edge devices to locally run 3B-7B LLMs/VLMs with an M.2 plug-and-play interface. Compatible with RK3588, RK3576, and other mainstream hosts. #1d1d1f #515154 #FFAE6C

Performance 20 TOPS NPU

https://www.forlinx.net/file.php?f=202608/f_695ee0648fd33037c44f46953ad36a36&t=webp&o=&s=&v=1787037479 https://www.forlinx.net/file.php?f=202608/f_e28fdeba513f63b5d14696918b7c5f80&t=webp&o=&s=&v=1787037352 #EEF8FE Smooth Offline Inference for 7B-Level Models Equipped with a 20 TOPS NPU supporting INT4/INT8/FP16 mixed precision. Runs 3B/7B LLMs offline (e.g., Qwen, LLaMA2) achieving 100+ tokens/s for smooth, low-latency edge AI without cloud dependency. #1d1d1f #515154 #EEF8FE

Versatility All-Scenario AI

https://www.forlinx.net/file.php?f=202608/f_55c88b1975ccd4bb45b67f287aa8c83b&t=webp&o=&s=&v=1787037665 https://www.forlinx.net/file.php?f=202608/f_8f94b76912428e3ade675df4bbe4f8c9&t=webp&o=&s=&v=1787037488 #F0FDF4 Extensive Model Compatibility Highly adaptable architecture supporting LLMs, VLMs, CNNs, Audio Models, and Object Detection. Meets diverse edge requirements across industrial inspection, smart security, and intelligent Q&A. #1d1d1f #515154 #CFE7FC

Memory Built-in DRAM

https://www.forlinx.net/file.php?f=202608/f_67e77e97669930930e755880a3337c12&t=webp&o=&s=&v=1787037673 #000000 3D Stacked High-Bandwidth Memory Integrates 2.5GB/5GB in-package DRAM, eliminating external DDR bottlenecks. Delivers hundreds of GB/s bandwidth to effectively overcome the "memory and bandwidth walls" for edge large models. #F5F5F7 #A1A1A6 #000000

Architecture Dedicated Co-Processor

https://www.forlinx.net/file.php?f=202608/f_b1f3f89dfeaa1237c78da4d1b749f6d4&t=webp&o=&s=&v=1787038013 https://www.forlinx.net/file.php?f=202608/f_3f5303d946c7f099629f3f86bad8c807&t=webp&o=&s=&v=1787038003 #FEF2F2 Decoupled Architecture for Dedicated AI Acts solely as an AI inference co-processor. Eliminates competition for host CPU, memory, or bandwidth, ensuring parallel operation and maximum system stability alongside host controllers like RK3588. #F5F5F7 #A1A1A6 #0A2543

Efficiency Cost-Effective

https://forlinx.net/file.php?f=202608/f_10bb30d5369cf00dcb10ea8897f1cc55&t=webp&o=&s=&v=1787038021 #000000 Low Power & High Cost-Effectiveness Delivers AI computing power comparable to the Jetson Orin Nano but with lower power consumption and superior cost-efficiency. The preferred choice for moderate-to-low power edge AI applications. #F5F5F7 #A1A1A6 #000000

Connectivity PCIe 2.1 1-Lane

https://forlinx.net/file.php?f=202608/f_8ecee7172757097bf618048c95591479&t=webp&o=&s=&v=1787038615 #E8EDF4 Flexible Expansion & Easy Upgrades Thanks to built-in DDR, the module has minimal reliance on external host bandwidth. It requires only a single PCIe 2.1 lane to fully unleash its AI computing capabilities. #1d1d1f #515154 #E8EDF4

Software Developer-Friendly

https://forlinx.net/file.php?f=202608/f_0fb856b005de589e7283e0e22a2013e8&t=webp&o=&s=&v=1787038621 #101941 Developer-Friendly Software Ecosystem Natively supports RKNN, TensorFlow, PyTorch, and ONNX. Includes a comprehensive RKNN SDK optimized for large models, enabling seamless migration and rapid algorithm deployment. #F5F5F7 #A1A1A6 #101941

Form Factor M.2 2280

https://forlinx.net/file.php?f=202608/f_319fb7928f942020dc3ff67d3f067f77&t=webp&o=&s=&v=1787038628 #000000 Standard M.2 Package for Rapid Integration Adopts a standard M.2 form factor for plug-and-play simplicity. Easily integrates with mainstream host platforms to instantly expand AI compute without complex hardware modifications. #F5F5F7 #A1A1A6 #000000

Support Comprehensive Docs

https://forlinx.net/file.php?f=202608/f_e985f27d5e9e5cb78356aea4839ea172&t=webp&o=&s=&v=1787039221 https://forlinx.net/file.php?f=202608/f_4cc2a93e159b0d56c70f831d6647a16e&t=webp&o=&s=&v=1787038639 #ECFDF5 Continuously Updated User Resources Access one-stop development materials including hardware manuals, carrier board schematics, compilation guides, and test routines to dramatically accelerate prototyping and mass production. #1d1d1f #515154 #EEEAF2

Applications Industry Ready

https://forlinx.net/file.php?f=202608/f_7c5b4c6f1ba8a7c9d3b91fbc7b45c440&t=webp&o=&s=&v=1787039230 https://forlinx.net/file.php?f=202608/f_d3bbeeb005209c276f1c99152a8414fa&t=webp&o=&s=&v=1787039213 #F8FAFC Wide Range of Industry Applications Ideal for Industrial IoT, Smart Power, Transportation, and Medical Terminals. Backed by full-process technical support including hardware adaptation and model migration for reliable deployment. #1d1d1f #515154 #F8FAFC

▊ Hardware Features

Hardware Parameters of AI Acceleration Module
Core Compute	Processor: RK1820 / RK1828
	RAM: Built-in 2.5GB / 5GB
	Compute Performance: 20 TOPS
	Computing Precision: INT4, INT8, INT16, FP8, FP16, BF16
Form Factor & Interface	Dimensions: M.2 2280
	Interface: M.2 M-Key
	Supported Host Controllers: RK3568, RK3576, RK3588, RK3572
Software & Ecosystem	Supported Operating Systems: Linux, Android
	Supported Models: Vision Models Audio Models Time-Series Prediction Models Large Language Models (LLM) Multimodal Models
	* Note: Supports full-featured Linux / Android drivers, and is adapted to mainstream AI frameworks.

▊ Documentation

Reference Manuals

Product Catalog

Hardware Related

Product Datasheet

User Guide

Carrier Board Schematic

Carrier Board PCB

SoM Pinmux

Software & BSP

OS Image

Testing Demo

Source Code

Manual

Request a Quote

Fill out the form below for volume pricing or technical specifications. We'll get back to you within 24 hours.

Name *

Email *

Company

Phone

Content *

I agree to the Privacy Policy

Reset

▊ Ordering Information

Purchase & Payment

1. Online Channels: We have an online store on Alibaba and DigiKey. Please contact us for details.

2. Direct Inquiry: For bulk orders, please email our sales team at [email protected].

3. Payment Method: Standard terms: 100% T/T (Telegraphic Transfer) in advance.

Shipping & Delivery

1. Dispatch Method: Orders are shipped via international express upon receipt of payment.

2. Lead Time: 5 working days for samples; 6 weeks for bulk orders.

3. Shipping Policy: Shipping and handling fees are the responsibility of the purchaser.

▊ Forlinx Recommended Modules & Accessories

Ara240 AI Acceleration Card

Model: FAI-ARA240-M
Interface: M.2
Features: High-efficiency 40 eTOPS M.2 accelerator for real-time generative AI and industrial edge inference.

Learn More

OK3588-C Single Board Computer

CPU: Rockchip RK3588/RK3588J
Architecture: 4x Cortex-A76+4x Cortex-A55
Frequency: [email protected]/1.6GHz, [email protected]/1.3GHz
RAM: 4GB/8GB/16GB LPDDR4x
ROM: 32GB/64GB/128GB eMMC

Learn More

OK3576-C Single Board Computer

CPU: Rockchip RK3576
Architecture: 4x Cortex-A72+4x Cortex-A53
Frequency: [email protected][email protected]
RAM: 2GB/4GB LPDDR4
ROM: 32GB eMMC

Learn More