When large AI models no longer rely on the cloud and compute sinks into every embedded device, Rockchip’s RK1820 coprocessor is turning “on-device intelligence” into reality.

The RK1820 is Rockchip’s add-on accelerator for flagship SoCs such as RK3576/RK3588. With advanced packaging, high performance, low power and multi-modal capability, it brings robust support for deploying big models at the edge.

RK1820 at a glance

Model support: up to 3 B/7 B-parameter on-device LLMs (16 K context)

Performance: > 100 tokens/s generation, < 0.1 s end-to-end latency

Modalities: text, voice, image, video; CNN compatible

Host links: PCIe 2.0 / USB 3.0; plug-and-play with RK3576, RK3588

Software: HuggingFace, PyTorch, GGUF; OpenAI-style API; C/Python bindings

Its high-bandwidth + low-power design breaks the energy/latency barrier for edge LLMs and delivers cloud-class responsiveness on site.

3-D TSV stack

Vertical die interconnect boosts bandwidth 10×, cuts power 30 % and halves footprint, integrating logic, memory and sensors in one package without adding volume.

About 3 B & 7 B models

3 B (30 B params): lightweight yet capable, runs on phones/edge boxes for chat, summarisation, coding, QA, translation, extraction—near 10 B-class quality offline.

7 B (70 B params): server/IPC/high-end laptop grade, handles long-doc summary, logical reasoning, code-gen, multi-turn dialogue, multimodal fusion—cloud-grade depth without the cloud.

Benchmarks

(see table in original)

Three key advantages

Host-ecology ready – PCIe/USB link needs no BSP change; works out-of-the-box on RK3568/RK3576/RK3588.

Partitioned compute – host runs OS/UI/I/O; RK1820 runs LLM, vision, semantics. Shared cache + high-speed bus isolate tasks and save power.

Independent upgrade – coprocessor evolves separately. Next-gen RK1860 will deliver > 64 TOPS and 13 B-model support at > 1 TB/s bandwidth, filling the domestic high-end gap.

Deployment snapshots

Education tablet – offline “AI teacher” with Qwen 3 B/7 B for spoken-English scoring, essay correction and tutoring without networking.

Auto cockpit – RK3588 + RK1820 supports 10+ concurrent voice agents, eliminating cloud latency for in-car multi-role dialogue.

Robotics – Qwen2.5-3 B emotion model on RK1820 gives low-power speech, sentiment and vision understanding.

Enterprise AI terminals – legacy RK3568/RK3399 boxes gain 100 token/s AI via USB/PCIe plug-in, adding ASR, image retrieval and text generation instantly.