Based on Rockchip AIoT SoC core boards and main boards. The table below drops straight into your selection report.
| Model | Process | CPU | NPU | video | Memory | Key interface | Typical application |
|---|---|---|---|---|---|---|---|
| RK3588 Flagship edge AI |
8nm | 4×Cortex-A76 @2.4GHz + 4×Cortex-A55 @1.8GHz | 6 TOPS (INT4/8/16/FP16) | 8K@60 dec / 8K@30 enc | LPDDR5 ≤32GB | 2×HDMI2.1, PCIe2.1×4, 4×SATA, dual GbE, 4×MIPI-CSI | Edge-AI server, on-device large model (7B/13B), robot, smart cockpit, industrial HMI RK3588S (compact) · RK3588J (industrial) · RK3588M (vehicle) |
| RK3576 Cost-effective flagship |
8nm | 4×Cortex-A72 @2.2GHz + 4×Cortex-A53 @1.8GHz | 6 TOPS (support Transformer) | 4K@120 dec / 4K@60 enc | LPDDR5 ≤32GB | HDMI2.1, PCIe, multi-display, USB/CAN/network | NVR, AI box, industrial HMI, AI POS, lightweight robot RK3576J (industrial) · RK3576M (vehicle) |
| RK3572 Mid-high-end AIoT |
— | 2×Cortex-A73 + 6×Cortex-A53 | 4 TOPS (sparse) | 8K@30 dec / 4K@30 enc | LPDDR4/4X/5/5X | HDMI2.1, PCIe2.1, multi-display, USB3.1 | Mid-high-end AIoT, commercial display, robot main controller RK3572M (vehicle) · RK3572V (vision 12M ISP) · RK3572S (cost down) |
| RK3568 Mainstream AIoT |
22nm | 4×Cortex-A55 @2.0GHz | 1 TOPS | 4K@60 dec / 1080P enc | LPDDR4x | USB3.0, PCIe3.0×2, SATA, dual GbE, 3×CAN | Industrial gateway, HMI, NVR, commercial display, self-service terminal RK3568J (industrial) · RK3568B2 |
| RK3566 Cost-effective |
22nm | 4×Cortex-A55 @1.8GHz | 1 TOPS | 4K@60 dec | LPDDR4x | USB2.0, MIPI, GMAC | Tablet, e-book, smart panel, lightweight terminal — |
| RK3562 entry-level AIoT |
22nm | 4×Cortex-A53 | 0.8 TOPS | 1080P | LPDDR4x | Streamlined general interface | Cost-sensitive entry-level AIoT device RK3562J(industrial) |
| RV1126 Smart vision |
— | 4×Cortex-A7 | 2 TOPS | AI-ISP 2.0 black-light full-color | — | MIPI-CSI, Ethernet | IPC, vision perception, behaviour-analysis camera RV1126K |
| RV1106/03 Lightweight vision |
— | Single Cortex-A7 | 0.5 TOPS | Lightweight vision | — | Streamlined | Entry-level AI camera, low-power vision node RV1106G2 / RV1103G1 |
| RV1126B Smart vision(upgrade) |
— | 4×Cortex-A53 | 3 TOPS | 4K@30 dec / 4K@45 enc · 12M AI-ISP | External DDR3/3L/4/LP3/4/4X | MIPI-CSI×2/4×2-Lane, RGMII+PHY, USB3.0 | Industrial IPC, vehicle vision, AI camera (RV1126 major upgrade) RV1126BJ (industrial -40~85) · RV1126BM (vehicle AEC-Q100) |
| RV1106B Entry-level vision |
— | Cortex-A7 + MCU | 0.5 TOPS | 4K@25 enc · 8M ISP | SiP DDR 512Mb/1Gb/2Gb | MIPI-CSI, RMII+PHY, USB2.0 | Battery IPC, entry-level AI camera, structured lightweight vision RV1106BG/BP |
| RK3308 audio / voice |
— | 4×Cortex-A35 | — | — | — | Audio codec, microphone array | Smart speaker, voice assistant, audio terminal RK3308M (vehicle) · RK3308J (industrial) |
| RK3506 Real-time control / HMI |
— | 3×Cortex-A7 + M0 | — | — | — | CAN/UART/SPI | HMI, motor control, PLC, gateway control panel RK3506J (industrial) |
| RK2118 Flagship audio (vehicle) |
— | 2×Cortex Star-M33 + 3×HiFi4 DSP | Audio-NPU 40 GOPS | — | 3MB SRAM + 64MB DDR | 8×SAI(>600 slot), PDM, SPDIF, 40ch ASRC, USB2.0/CAN | Vehicle audio amp, Soundbar, Dolby Atmos/DTS-X, smart speaker RK2118M/M2(vehicle AEC-Q100) |
| RK2116 RISC-V audio |
— | RISC-V MCU + 2×HiFi4 DSP | — | — | 1.28MB SRAM | 8×SAI, PDM, 24ch ASRC, 4AD/4DA Codec, USB2.0/CAN | Vehicle / consumer audio, voice frontend RK2116M/M2(vehicle AEC-Q100) |
| RK2108D Low-power AI voice |
— | Cortex-M4F @396MHz + HiFi3 DSP @594MHz | — | MIPI/RGB ≤720P | 1MB Share + BootRom | I2S×2 / PDM 6ch / Codec ADC 2ch / USB2.0 / SDIO3.0 | AI voice interaction, low-power IoT audio (M08D Module main controller) M08D Module (ours) |
Pairs with RK3576 / RK3588: the main controller runs business & tooling while the NPU card specializes in LLM/VLM inference (3D-stacked DRAM, the only on-device bandwidth of several hundred GB/s).
| Model | NPU compute | Large-model capability | On-board DRAM | Interface | Status |
|---|---|---|---|---|---|
| RK1828 | 20 TOPS | 3B LLM/VLM(Qwen3 family) | 5GB 3D DRAM | PCIe2.1 / RGMII / USB3.1 | mass production |
| RK1820 | 20 TOPS | 7B LLM/VLM | 2.5GB 3D DRAM | PCIe2.1 / RGMII / USB3.1 | mass production |
| RK186X | 64 TOPS | 7B / 13B LLM/VLM | 2.5/5/10GB | PCIe3 / USB3.0 / C2C | ES 2026Q3 |
Measured Qwen3-1.7B(RK1828, W4A16) : Prefill 2387 TPS / Decode 138 TPS, far exceeding RK3588(298/13.9). Note: the official roadmap (2026-05) does not list "RK1280"; the AI NPU co-processor line is RK1820/1828/186X — confirm the specific model with Rockchip sales.
# 1) Install RKNN-Toolkit2 (Python 3.8+)
pip install rknn-toolkit2
# 2) Quantize and export the RK3588 inference model
from rknn.api import RKNN
rknn = RKNN()
rknn.config(target_platform='rk3588')
rknn.load_onnx(model='model.onnx')
rknn.build(do_quantization=True, dataset='./calib.txt')
rknn.export_rknn('model.rknn')
# 3) On-board inference (RK3588, C/Python API)
rknn.init_runtime(target='rk3588')
outputs = rknn.inference(inputs=[img])
Full flow in the Technical Docs / SDK guide. BesTom provides schematics and board-level support packages.