RK3588 Compared: A Practical Guide for Edge AI Buyers

iTOP-RK3588 Single Board Computer

The Rockchip RK3588 has become a reference point for edge AI single-board computers, largely because it pairs an eight-core CPU with a 6 TOPS neural processing unit at a price far below most GPU-based platforms. That combination has made "rk3588 vs raspberry pi 5" and "best SBC with NPU" two of the most searched comparison phrases in this category. This guide compares the platforms on published specifications, explains what the NPU can realistically do, and points to the TOPEET iTOP boards that implement it.

Every figure below is a published specification, not a benchmark. Where vendors quote NPU throughput differently, the basis is stated. For the fully documented TOPEET board, shop-listing figures and packed-datasheet figures are both shown where they disagree.

How the three platforms compare on paper

Specification RK3588 SBC Raspberry Pi 5 Jetson Orin Nano Super
CPU 4× Cortex-A76 + 4× Cortex-A55, octa-core 4× Cortex-A76 (BCM2712) 6-core ARM CPU
NPU / AI 6 TOPS on-chip (INT8) None on board Up to 67 TOPS (GPU/accelerator)
Max memory Up to 32 GB Up to 16 GB 8 GB module
PCIe PCIe 3.0 ×4 PCIe 2.0 ×1 PCIe 4.0
Ethernet Dual Gigabit Ethernet on industrial designs Single Gigabit Ethernet Gigabit Ethernet
Video decode 8K@60fps 4K 4K-class
Typical use Edge AI, vision, signage, industrial Education, hobby, light desktop High-end AI and robotics

The pattern is clear: the Raspberry Pi 5 is a strong general-purpose and education board, but it has no on-board NPU, a single PCIe lane, a single Gigabit port, and a 16 GB ceiling. The Jetson Orin Nano Super leads on raw AI throughput but at a much higher cost and with a steeper software stack. The RK3588 sits between them on both price and capability, which is why it appears so often in "raspberry pi alternative" searches for industrial and commercial projects.

What 6 TOPS can actually run

6 TOPS is an INT8 inference figure. In practice it is enough for real-time object detection with YOLO-class models, face detection across several camera streams, speech recognition, and lightweight on-device language models. It is not a substitute for a data-centre GPU, and it will not run large models at interactive speed without quantisation and an accelerator.

Two details decide whether a board is usable in production. The first is memory bandwidth, which constrains how quickly a model can be fed. The second is the software toolchain: RKNN-Toolkit2 for converting and deploying models, and RKLLM for language models. A board with strong silicon but no maintained toolchain will cost more in engineering time than it saves in unit price.

Multi-display and video: the other reason buyers choose RK3588

Beyond AI, the RK3588 integrates an 8K video decoder and multiple display interfaces. Industrial carriers bring out HDMI, eDP, DisplayPort and MIPI-DSI, allowing several independent outputs from one board. For digital signage, kiosks and HMIs this removes the need for separate graphics hardware and lowers total system cost.

TOPEET implementation: the iTOP-RK3588 family

The iTOP-RK3588 single board computer is sold as a carrier plus system-on-module, with commercial and industrial memory grades, connector and stamp-hole form factors, and optional 4G or 5G. The 4G and 5G kits are listed separately, and a compact iTOP-RK3588S variant is available for space-constrained designs. For production designs that want to separate compute from carrier, the iTOP-RK3588 connector SoM is the module-only option.

For projects that need more AI headroom than the on-chip NPU, the RK1828 / RK1820 M.2 accelerator card adds inference capacity over M.2, with the software stack supplied by the same vendor ecosystem.

Where the RK3568 and RK3576 fit

Not every project needs flagship performance. The iTOP-RK3568 platform is a mature, cost-effective choice for HMIs, control panels and entry-level signage; its documented kit uses a 190 mm × 125 mm carrier with a 65 mm × 55 mm module on four 80-pin connectors, with commercial (0 °C to +70 °C) and industrial (−40 °C to +85 °C) grades. The iTOP-RK3576 sits between the two, with a 6 TOPS NPU and strong video capability at a lower cost than RK3588. Choosing the smaller part avoids the over-specification that inflates BOM cost and supply risk.

Which board should you choose

  • Learning, prototyping, light desktop: Raspberry Pi 5 remains the lowest-friction option with the largest community.
  • Maximum AI throughput and an established NVIDIA stack: Jetson Orin Nano Super.
  • Vision, signage, gateways and industrial products at scale: RK3588, where NPU, PCIe 3.0, dual Gigabit Ethernet and 8K decode justify the platform.
  • Cost-sensitive HMI and control with long supply: RK3568 or RK3576.

Frequently asked questions

Is RK3588 better than Raspberry Pi 5? For AI and industrial I/O, yes: it has a 6 TOPS NPU, PCIe 3.0 ×4, dual Gigabit Ethernet options and up to 32 GB. For learning and general hobby use, the Pi's ecosystem is still easier.

Can RK3588 run ChatGPT or other LLMs locally? Small quantized models can run on-device, especially with an M.2 accelerator. Large models require cloud or a much bigger GPU.

Does the NPU work with PyTorch models? Models are normally converted to RKNN format with RKNN-Toolkit2; not every PyTorch operation converts cleanly, so check model compatibility early.

Where can I get the full specifications? Each linked product page lists its options; for the iTOP-RK3568 the packed datasheet is the authoritative source for dimensions and temperature grades.

Related reading: How to choose an industrial SBC or SoM for production equipment and our edge AI computer buying guide.

0 comments

Leave a comment

Please note, comments need to be approved before they are published.