AI Potluck
Infrastructure / Edge AI hardware

Rockchip RK3588

Rockchip

Application-processor SoC from Fuzhou Rockchip rather than a board. It is an octa-core Arm design pairing four Cortex-A76 with four Cortex-A55, a Mali-G610 MC4 GPU and a 6 TOPS triple-core NPU supporting INT4, INT8, INT16, FP16, BF16 and TF32, with up to 32GB of LPDDR4X or LPDDR5 and PCIe 3.0. It is the silicon behind boards including Orange Pi 5, Radxa Rock 5B and Khadas Edge2.

Documentation is unusually complete for this class, with a public technical reference manual and datasheet, while the SDK ships as prebuilt binaries and the firmware stays closed. Verified 2026-08-13 via the Rockchip brief datasheet, the product page, the rknn-toolkit2 repository and the rkbin repository.

Openness

3 high confidence
3.0
toolchain
partial (NPU SDK on GitHub but proprietary RKNN license + prebuilt blobs, rknpu kernel driver open)
datasheets
public (full TRM + datasheet public, no NDA)
blobs
required (proprietary GPU/NPU firmware)

Documentation is unusually open (full TRM and datasheet public) and the RKNN toolkit is openly distributed, but the SDK ships as prebuilt binaries under a proprietary license with closed firmware. The brief datasheet downloads straight from rock-chips.com with no registration or NDA, the RKNN SDK licence in airockchip/rknn-toolkit2 is the proprietary "RKNN SDK License" (no reverse engineering, use restricted to Rockchip products), and rkbin distributes the DDR/BL31/OPTEE firmware as prebuilt binaries.

Adoption

5 high confidence
5.0

Powers a very large multi-vendor set of SBCs and modules (Orange Pi, Radxa, Khadas, FriendlyELEC, Firefly), widely available at retail and broadly supported in community Linux. The Armbian community comparison tabulates five RK3588/RK3588S boards from five different vendors, three of which (orange-pi-5, radxa-rock-5b, khadas-edge2) are separately on this map and separately confirmed on sale.

Capability

4 high confidence
4.0

The 6 TOPS NPU is genuine edge-AI capability and the RKLLM SDK runs quantized LLMs on-chip, but modest versus dedicated accelerators. The RKLLM repository ships an RK3588-series runtime and toolkit with working w8a8 demos for Qwen2.5-VL-3B, Qwen3-VL-2B, InternVL3-1B and DeepSeek-OCR, and Rockchip's own product page reads 6 TOPS triple-core NPU. Generative workloads actually running on-chip is what keeps this a step above the 6 TOPS boards built on it.

Verified 2026-08-13