Skip to main content

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​

ParameterValue
ArchitectureOcta-core 64-bit (4×Cortex-A72 + 4×Cortex-A53) + Mali-G52 MC3 + in-house NPU
Process NodeSecond-generation 8nm
FP16 / BF16 ComputeSupports FP16/BF16 (NPU mixed precision; specific value not broken out)
INT8 Compute6 TOPS (INT8; supports sparse acceleration)
FP32 ComputeNot disclosed
Memory CapacityDual-channel 32-bit LPDDR4/LPDDR4x-4266 / LPDDR5-4800
Memory TypeLPDDR4 / LPDDR4X / LPDDR5
Memory BandwidthNot disclosed
TDPAbout 3.2 W at full NPU load (lower for the whole device)
InterconnectPCIe 2.1/SATA3, USB3.0, dual RGMII, CAN FD, MIPI CSI/DSI, HDMI2.1/eDP/DP
InterfaceSoC on-board (core board / development board)
Launch2024-09 (mass production)
Mass Production/AvailabilityMass 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​

ParameterValue
CompanyRockchip Electronics Co., Ltd. (Rockchip)
HeadquartersFuzhou, China
Founded2001

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)

References​