Rockchip RK3568
The industrial workhorse of the range. Quad Cortex-A55, 4K video, triple display and — uniquely in this class — native dual Gigabit Ethernet. If you need two network segments or three screens without stepping up to the flagship, this is the chip.
Boards using it: KTM-S58, KTM-S68, KTM-MS68.
At a glance
| Process | 22 nm |
| CPU | Quad-core Cortex-A55, up to 2.0 GHz |
| GPU | Mali-G52 2EE — OpenGL ES 1.1/2.0/3.2, OpenCL 2.0, Vulkan 1.1 |
| NPU | ~1 TOPS, INT8/INT16 hybrid |
| Video decode | 4K@60 H.265/H.264/VP9 |
| Video encode | 1080P@60 H.265/H.264 |
| ISP | 8 MP @ 30 fps, HDR |
| Displays | Up to 3 (two simultaneous on most boards) |
| Ethernet | Dual Gigabit (native) |
| Memory | LPDDR3/4/4X, DDR3/4 |
What it means for your product
Video playback
The RK3568 decodes 4K at 60 fps in hardware for H.265, H.264 and VP9. For signage playing 4K content this is comfortable headroom — the CPU stays free for your application while the video engine does the work.
It does not decode 8K. If your content pipeline is 8K, you need RK3588. For the overwhelming majority of signage — 1080p and 4K — the RK3568 is the right tool and runs cooler doing it.
Encoding is 1080p, not 4K. This matters only if your product records or streams video out; for playback-only signage it is irrelevant.
Displays
Three display interfaces (from HDMI 2.0, eDP, MIPI-DSI, LVDS), with two able to run simultaneously on most KTM boards. Independent content on each — not a mirror. See the board page for which physical connectors are populated.
On-device AI
The ~1 TOPS NPU handles lightweight inference — a single-camera face match, basic object or people detection, presence sensing. It supports INT8/INT16 and converts models from TensorFlow, PyTorch, Caffe and ONNX through Rockchip's toolchain.
Be realistic about the ceiling. Multi-camera analytics or a large model will saturate it. If AI is central to the product, benchmark your model on the board early, or move up to RK3588's 6 TOPS. If AI is a feature rather than the point, 1 TOPS is enough.
Why dual Ethernet is the headline feature
The RK3568's distinguishing trait in this class is native dual Gigabit Ethernet. That lets one device sit on two network segments at once — a payment network and a content network, or an isolated device VLAN plus an uplink. The cheaper RK3566 has one. If your deployment needs network separation, this is the reason to choose RK3568 over RK3566.
RK3568 vs RK3566
They are close relatives — same CPU family, same GPU class, same 4K decode. The RK3568 adds what industrial deployments need:
| RK3568 | RK3566 | |
|---|---|---|
| Ethernet | Dual GbE | Single |
| PCIe | PCIe 3.0 ×2 | PCIe 2.1 |
| Displays | Up to 3 | Up to 2 |
| Package / IO | Larger, more IOMUX | Smaller |
| Cost | Higher | Lower |
Choose RK3566 for cost-sensitive single-screen products; RK3568 when you need the second network, the third display, or the extra I/O.
Software
Runs the same firmware and API as every KTM board. Power scheduling, GPIO, display control and network configuration are identical — see the ItempApi reference.
For on-device AI, Rockchip's RKNN toolchain converts and deploys models to the NPU. See the RKNN deployment guide for the full workflow.
Related
- Boards: KTM-S58, KTM-S68, KTM-MS68
- Step up: RK3588 for 8K and 6 TOPS AI
- Step down: RK3566 for single-display, cost-sensitive builds
- All SoCs
Chip capability figures are from Rockchip published specifications and may vary by firmware and board configuration. Board-specific values are on each board page.