Rockchip RK3576
Product Overview
The RK3576 is the second-generation 8nm high-performance, low-power AIoT application processor SoC launched by Rockchip in 2024 (mass production in September 2024). It adopts an octa-core 64-bit architecture (4×Cortex-A72 @2.2GHz + 4×Cortex-A53 @1.8GHz), integrating a Mali-G52 MC3 GPU and a 6 TOPS in-house NPU — the same NPU compute specification as the flagship RK3588 — but positions itself for mid-to-high-end edge devices with better energy efficiency and scenario-oriented design.
The RK3576 supports 8K@30fps decoding, 4K@60fps encoding, and a 16MP ISP. The NPU supports INT4/INT8/INT16/FP16/BF16/TF32 mixed precision and is specially optimized for 2B-parameter multimodal models (such as Qwen2-VL-2B), reaching token generation speeds of 10+ per second. Its NPU draws only about 3.2W at full load, about 22% lower than the RK3588, making it better suited to battery-powered and mobile endpoint devices.
Core Specifications
| Parameter | Value |
|---|---|
| Architecture | Octa-core 64-bit (4×Cortex-A72 + 4×Cortex-A53) + Mali-G52 MC3 + in-house NPU |
| Process Node | Second-generation 8nm |
| FP16 / BF16 Compute | Supports FP16/BF16 (NPU mixed precision; specific value not broken out) |
| INT8 Compute | 6 TOPS (INT8; supports sparse acceleration) |
| FP32 Compute | Not disclosed |
| Memory Capacity | Dual-channel 32-bit LPDDR4/LPDDR4x-4266 / LPDDR5-4800 |
| Memory Type | LPDDR4 / LPDDR4X / LPDDR5 |
| Memory Bandwidth | Not disclosed |
| TDP | About 3.2 W at full NPU load (lower for the whole device) |
| Interconnect | PCIe 2.1/SATA3, USB3.0, dual RGMII, CAN FD, MIPI CSI/DSI, HDMI2.1/eDP/DP |
| Interface | SoC on-board (core board / development board) |
| Launch | 2024-09 (mass production) |
| Mass Production/Availability | Mass production in 2024-09 |
Key Features
- 6 TOPS in-house NPU: supports INT4/INT8/INT16/FP16/BF16/TF32, with a dynamic sparse acceleration engine.
- Mid-range energy-efficiency benchmark: only 3.2W at full NPU load, about 22% lower than the RK3588.
- 8K multimedia: 8K@30fps decoding, 4K@60fps encoding, multi-screen heterogeneous display (different sources).
- 16MP ISP: HDR (up to 120dB), 3DNR, multiple MIPI CSI inputs.
- 2B model optimization: dedicated acceleration for lightweight large models such as Qwen2-VL-2B.
- Rich interfaces: PCIe/SATA/USB3.0/CAN FD, standard Android/Linux SDK, domestic OS support.
Vendor Information
| Parameter | Value |
|---|---|
| Company | Rockchip Electronics Co., Ltd. (Rockchip) |
| Headquarters | Fuzhou, China |
| Founded | 2001 |
Use Cases
- ✅ Mid-range AIoT, edge computing, industrial control, smart cockpits, service robots, battery-powered terminals, conference whiteboards
- ❌ Extremely high-density video analytics (weaker than dedicated AI accelerator chips)
Related Cards
- /docs/cards/others/rockchip-rk3588 — Series flagship (same compute)
- /docs/cards/others/hailo — International edge NPU comparison
- /docs/cards/others/moore-threads-mtt-s5000 — Domestic GPU compute card comparison