Rockchip RK3588 and RK3588S — flagship Cortex-A76 edge AI SoC

The Rockchip RK3588 is the current flagship Chinese application processor and the reference point for edge AI on a budget. Four Cortex-A76 cores at 2.4 GHz alongside four Cortex-A55 cores at 1.8 GHz, a Mali-G610 MP4 GPU, a 6 TOPS NPU, 8K H.265 decode, dual 2.5-or-1 GbE plus 10 GbE-PHY-friendly SerDes, PCIe Gen3 ×4, dual USB 3.1 Gen 1, and the ability to drive multiple 4K displays from one chip. It is the SoC behind every recent high-end Chinese SBC and a steadily growing fraction of the industrial edge-AI deployment base, where its price-versus-capability point has displaced NXP i.MX 8M Plus and put real pressure on the NVIDIA Jetson Orin Nano. This page covers the silicon details, the RK3588 versus RK3588S split, board landscape, RKNN Toolkit, and the comparison with Jetson Orin.

The RK3588 versus RK3588S split

Both parts share the same silicon die but differ in package and peripheral exposure.

The RK3588 is the choice for designs needing NVMe SSD via PCIe, multiple network interfaces, or multi-monitor output. The RK3588S is the cheaper, smaller-package variant for products where PCIe ×4 is overkill.

Boards

The Orange Pi 5 Plus and Radxa ROCK 5B are the most common targets for general-purpose RK3588 software work. The NanoPi R6S has become a popular OpenWrt router target due to the dual 2.5 GbE configuration.

RKNN Toolkit and the NPU

The 6 TOPS NPU is the differentiator for RK3588 against earlier Rockchip parts. It is implemented as three independent NPU cores, each delivering 2 TOPS, gangable for a single large model or split across three concurrent inference jobs. Supported numerics are INT4, INT8, and FP16. The flow goes through the RKNN Toolkit: train a model in PyTorch/TensorFlow/Caffe, export to ONNX, run rknn-toolkit2's converter to produce a quantised .rknn file (the converter handles INT8 quantisation with a calibration dataset, layer fusion, and weight repacking), then deploy and run via rknn-toolkit-lite2 (Python) or the C API.

Published performance numbers from Rockchip's RK3588 benchmarks:

LLM inference uses the NPU for matrix multiplies plus CPU for token sampling and attention bookkeeping. Performance trails Jetson Orin Nano on transformer-heavy workloads but is workable for chatbot products with modest latency targets.

import numpy as np
from rknnlite.api import RKNNLite
 
# Load a pre-converted RKNN model and run inference
rknn = RKNNLite()
ret = rknn.load_rknn('./yolov8n.rknn')
ret = rknn.init_runtime(core_mask=RKNNLite.NPU_CORE_0_1_2)  # use all three NPU cores
 
img = np.random.randint(0, 255, (1, 640, 640, 3), dtype=np.uint8)
outputs = rknn.inference(inputs=[img])
 
# outputs is a list of np.ndarrays per model output head
for i, out in enumerate(outputs):
    print(f'output {i}: shape={out.shape}, dtype={out.dtype}')
 
rknn.release()

Comparison with NVIDIA Jetson Orin Nano

This is the comparison most engineers actually want to see.

RK3588 is roughly half the price of Jetson Orin Nano 4 GB at similar total capability, but the software ecosystem is meaningfully behind. For products where the engineer can bring up RKNN on a custom model, RK3588 is the cost winner. For products depending on prebuilt TensorRT models, CUDA-only libraries, or NVIDIA's DeepStream frameworks, Jetson is the answer. A practical heuristic: a single well-defined inference job (object detection on one camera, keyword spotting) suits RK3588 well; heterogeneous model stacks or anything pulling from NVIDIA NGC favours Jetson.

Software stack and kernel status

The RK3588 software situation is more painful than the RK3566/68 because mainline support lags. As of mid-2026:

Practical recommendation for production: vendor BSP kernel 5.10 plus Armbian or Debian userspace; plan the rebase to upstream 6.x over 2026-2027 if product lifetime warrants.

Use cases

• Edge AI workstation — RK3588 with 16 or 32 GB RAM and NVMe SSD as a developer's edge inference target.
• Multi-camera analytics — four IP cameras decoded via rkvdec, YOLOv8n inference on each stream via NPU partitioning.
• Robotics CPU — high-level navigation and perception on RK3588 paired with a Cortex-M for motor control (similar pattern to TI Sitara AM62).
• Mini desktop — full Linux desktop replacing a low-end x86 office machine.
• 8K signage — RK3588 driving 8K H.265 streams for premium digital signage.
• Local LLM appliance — RK3588 with 32 GB RAM running quantised 7B-13B language models for on-premises chat.

Cross-references

• Rockchip landing — vendor overview.
• Rockchip RK3568 — lower-cost sibling.
• MediaTek Genio 1200 — direct competitor.
• Cortex-A76 — core architecture details.
• Cortex-A55 — little-cluster core details.
• rockchip.com.cn RK3588 page (external).
• rockchip-linux/rknn-toolkit2 on GitHub (external).
• Orange Pi 5 Plus product page (external).
• Radxa ROCK 5B documentation (external).
• Armbian for RK3588 (external).