Direct answer
Choose RK3588 when the product itself needs the larger platform: several camera paths, demanding video, high-resolution or multiple displays, and substantial expansion I/O. Choose RK3588S when you need the RK3588-class compute and media capability but the final product can live within the smaller package’s interface budget. Choose RK3568 when the product’s camera, video, display and local-AI requirements are lighter and its software and thermal envelope suit a quad-core Cortex-A55 platform.
That is a capability decision, not a race. An RK3588S is not a deliberately slower RK3588; the important question is which interfaces the selected SoC, module and carrier board actually make available to the finished product. Likewise, a higher TOPS figure does not prove that a particular model, camera pipeline or thermally enclosed unit will meet its target.
Comparison at a glance
| Product requirement | RK3588 | RK3588S | RK3568 |
|---|---|---|---|
| Compute and media class | Flagship octa-core RK3588 class with a 6-TOPS NPU and 8K-class media/display capability. | RK3588-class CPU, GPU, NPU and media capability in a package intended for designs that can accept a reduced external-interface budget. | Quad-core Cortex-A55 class with a much smaller NPU; suitable where lighter local inference and 4K-class video are enough. |
| Camera and video | Start here for multi-camera vision, high-resolution capture, video analytics, NVR and demanding video/display systems. | Assess sensor count, CSI lanes, HDMI input and board routing before choosing it for the same class of vision work. | Fits conventional HMI, gateway, industrial-control and lighter edge-vision products. Validate the exact sensor and encoder requirements. |
| Displays and expansion | Best fit when the product genuinely needs rich display options plus PCIe, SATA, USB, Ethernet or other high-speed expansion. | Good fit only when the product can work with the package and module’s exposed interfaces; check the carrier schematic, not a generic RK3588 feature list. | Useful for more modest display and I/O needs, where the product architecture does not depend on the flagship platform’s breadth. |
| What to validate first | Memory configuration, heat path, camera topology, display routing and the selected module/carrier. | The exact package pinout and the module/carrier’s exposed camera, display, storage and expansion interfaces. | Model compatibility and latency, camera/video format, memory, operating system support and the final thermal envelope. |
Table note: SoC capability is not the same as board availability. A module or carrier can omit interfaces that exist in the silicon.
What changes between RK3588 and RK3588S?
The useful distinction is package and external connectivity, not a simplistic performance ranking. RK3588 is the fuller platform for products that need to bring many high-speed functions out to the board. RK3588S is the alternative for designs that want the same general CPU, GPU, NPU and media class but do not need every external lane or interface.
Before selecting either part, make a pin-and-interface list for the finished product: cameras and lane counts, display outputs and resolutions, storage, PCIe devices, USB, Ethernet, audio and any special industrial interfaces. Then compare that list to the exact Rockchip package documentation and to the module or carrier board schematic. A datasheet feature that is not routed on the chosen board is not available to the product.
Local AI: start with the model, not TOPS
RK3588 is commonly described as a 6-TOPS NPU platform, while RK3568 is a substantially smaller NPU class. That difference matters for the size and number of models you can run, the precision you use, the camera resolution and frame rate you can sustain, and the amount of headroom left when the device is hot or doing other work. It does not, by itself, state an inference latency or guarantee a result for an LLM, detector or vision-language model.
For an AI feature, define the actual job first: model format, input size, quantisation, target latency, number of simultaneous streams, memory budget and acceptable accuracy. Convert and test the model with the RKNN toolchain on the intended operating system and board. Test with representative camera frames or production data, not only a vendor demo. This is also where memory bandwidth, CPU-side pre-processing, post-processing and cooling often become visible constraints.
For a small gateway that classifies a modest sensor or image stream, RK3568 may be adequate. For several live video streams or a heavier vision pipeline, the RK3588 class is often the more appropriate starting point. The decision should be proven by the workload, not inferred from a peak operations-per-second number.
Camera and video selection
Write the camera requirement as an architecture: number of sensors, resolution and frame rate of each, CSI-2 lane allocation, ISP requirements, HDR needs, codec direction, simultaneous streams and display output. Then assess the silicon, the module and the board separately.
RK3588 is designed for multi-camera inputs, an advanced ISP, 8K-class decode/encode and multi-display capability. It is the natural candidate for NVR, multi-camera vision, video analytics and display-heavy equipment when the whole system needs those features. RK3568 supports 4K-class decode, 1080p-class encode and an 8 MP ISP according to Rockchip’s published material, which can be a better fit for a more conventional HMI, gateway or lighter vision design.
RK3588S can preserve the RK3588-class compute/media choice where the physical interface plan fits its package and board. Do not assume the word “camera” or “8K” on an SoC comparison page settles the product decision: confirm sensor support, lanes, drivers, image-processing pipeline and the exact simultaneous-use case.
Display, storage and expansion I/O
Display, storage and expansion are frequently the deciding constraints. RK3588 lists eDP, DP, HDMI 2.1 and MIPI display interfaces, high-speed PCIe, SATA, USB and Ethernet options. That breadth is valuable only when the design needs it and the board routes it. It can be unnecessary when a single display, ordinary storage and a compact interface set fully cover the product.
With RK3588S, treat the module/carrier selection as part of the chip decision. Check which display outputs, camera connectors, PCIe lanes, USB ports, Ethernet options and storage interfaces are physically exposed. With RK3568, check the same things, while being especially clear about video, display and NPU limits. A well-matched carrier is more useful than an unmatched SoC with unconnected capabilities.
Practical product fits
- RK3588: multi-camera recorders, vision inspection systems, high-resolution display appliances, edge-AI systems with substantial local video work, and products that need broad high-speed expansion.
- RK3588S: compact products that need the RK3588 class of compute or media capability but have a defined, narrower external-interface plan.
- RK3568: industrial HMIs, gateways, control panels, cloud terminals and lighter edge-AI or edge-vision products where its video, display and local-inference envelope is sufficient.
Implementation limits to resolve before committing
- Lock the intended module or carrier, then verify its schematic and supported software image against the product’s real interfaces.
- Run the selected AI model through the RKNN toolchain and measure it with the intended input, resolution and stream count.
- Validate the actual camera sensors, codec path, displays, storage and peripherals together. Individual feature claims do not prove a simultaneous configuration.
- Measure thermals in the proposed enclosure and power mode. Sustained local AI and video performance depends on the heat path, memory choice and system load.
FAQ
Is RK3588S slower than RK3588?
Not as a general rule. The meaningful difference is the package/interface trade-off. Confirm the exact CPU, GPU, NPU, media and external interfaces required by the selected design, module and carrier.
Which chip is best for local AI?
There is no universal answer. RK3588-class hardware gives more NPU headroom for demanding local workloads, while RK3568 can suit lighter workloads. Test the actual model, precision, input size and concurrent streams with RKNN on the target board.
Can I choose from the SoC datasheet alone?
No. The board may expose fewer interfaces than the SoC, and system performance depends on memory, software, sensors, peripheral routing and thermal design. Use the datasheet to form a shortlist, then validate the intended implementation.
Sources
Feature summaries should be checked against the current primary documentation for the exact revision and package: Rockchip RK3588 product page; Rockchip RK3568 datasheet; and the Rockchip download centre. Confirm current package-specific material directly with Rockchip or the module supplier before release.