Overview
The RK3588 is Rockchip's flagship embedded processor and the controller most Chinese-made compute boxes are built on. An 8nm process, an octa-core layout, an in-house 6 TOPS NPU, full multimedia capability and extensive I/O expansion.
Its position is the all-rounder: AI inference, 8K encode and decode, multi-display output and peripheral integration on a single chip, for scenarios where one device has to do several jobs.
Key Features
- Octa-core plus a 6 TOPS NPU: Four Cortex-A76 cores alongside four Cortex-A55 cores, with a Mali-G610 MC4 GPU. The NPU uses a three-core in-house design rated at 6 TOPS INT8, with cores that can run in parallel or independently, supporting mixed INT4, INT8, INT16, FP16, BF16 and TF32 quantisation.
- 8K encode and decode: Hardware decode for 8K@60fps H.265, VP9 and AVS2, decode for 8K@30fps H.264, and hardware encode up to 8K@30fps. It can run 32 simultaneous 1080P@30fps decode streams or 16 1080P@30fps encode streams.
- Multi-display and multi-camera: Dual HDMI 2.1 with 8K@60Hz on a single output, plus DP 1.4, eDP and MIPI-DSI for four independent displays natively. An integrated 48MP ISP supports up to six simultaneous MIPI camera inputs.
- Toolchain and ecosystem: RKNN Toolkit2 converts and quantises models from PyTorch, TensorFlow, Caffe and ONNX. Ubuntu, Android and Debian are supported, with native Docker support.
- Measured performance: With 8GB of LPDDR4X, YOLOv8s reaches 75 to 82 FPS at 640×640 input, YOLOv8x around 32 to 38 FPS and ResNet50 over 210 FPS. Inference is hardware-isolated and does not compete with CPU or GPU resources.
- Power and cost: A complete box draws around 5W idle and about 12W at full load, quiet under passive cooling. Complete units in 2026 sit roughly in the CNY 600 to 1,200 range, among the cheaper options at this compute level.
Use Cases
- Smart security with multiple video feeds and real-time object detection
- Industrial vision for defect detection and product sorting
- Robotics platforms combining multiple cameras and voice modules on low power for longer runtime
- Multimedia terminals needing 8K encode and decode with multiple display outputs
Pros
- CPU, NPU and GPU work heterogeneously, covering AI and multimedia together
- One-click conversion from mainstream frameworks through RKNN, with friendly quantisation
- Complete I/O set including Ethernet, RS485, CAN, PCIe and MIPI
- Low system cost, passive cooling and a large body of community material
Pricing
Sold as chips and development boards. In 2026 complete RK3588 boxes sit roughly in the CNY 600 to 1,200 range, consumer boards such as the Orange Pi 5 Max around CNY 600, and industrial models above CNY 1,000.
Summary
The RK3588 wins on breadth and price. A 6 TOPS NPU handles the YOLO family and light vision models comfortably, while 8K codecs and abundant I/O make it easy to deploy in industrial and security settings. The limits are equally clear: 6 TOPS suits roughly 1B-class models, and on-device large model work needs an add-on compute module or a different platform. Confirm model size first; for pure vision tasks it is a strong value answer.
Version History
- RK3588 with RK1828 compute module (2026): A 16 TOPS M.2 co-processing module plugs into RK3588 boards over PCIe, splitting control and heavy inference
- RK3588 launch (2022-09): 8nm octa-core design with an in-house 6 TOPS NPU and 8K encode and decode