Straight answer first: this is not a question of whether RK3576 can replace RK3588, but of where it saves money and where it costs you. If the product needs 8K, high-bandwidth expansion or large local models, RK3588 has no substitute today. If it needs steady multi-channel vision inference, an HMI or a retail motherboard, RK3576 is usually the smarter tier.
1. What actually went up
- Silicon: RK3588 has moved roughly 25–30% across two visible rounds, with further pressure expected through early 2026 — your quotation is the authority.
- Memory: tight DDR supply makes LPDDR4/5 the least predictable line in the BOM, and it decides whether your target price survives.
- Hidden: unclear lead times turn into inventory exposure, locked capital and re-qualification of substitute parts.
2. Where the four platforms actually sit
| Platform | CPU | NPU | Display & video | Memory | Typical product | Note |
|---|---|---|---|---|---|---|
| RK3588 | 4×A76 + 4×A55, up to 2.4GHz | 6 TOPS | 8K60 decode, HDMI 2.1 out, multi-display | LPDDR4/4X/5 | AI boxes, multi-screen signage, vision hosts | Highest price and supply pressure |
| RK3576 | 4×A72 + 4×A53, up to 2.2GHz | 6 TOPS | 4K120 decode & display, three screens | LPDDR4/4X/5 | Mid-tier AI boxes, industrial HMI, retail boards | Best value when 8K-class needs are absent |
| RK3572 | 2×A73 + 6×A53, 2.2GHz | 4 TOPS | Dual display 4K60 + 2K60 | LPDDR4/4X | Dual-screen vending, access control (2026) | Too small for heavy model inference |
| RK3568 / RK3566 | 4×A55, up to 2.0GHz | 0.8–1 TOPS | 4K60 | LPDDR4 | Lite gateways, basic HMI | Not enough headroom for real AI |
Disclaimer: specifications are compiled from public datasheets and third-party benchmarks; always validate against the vendor datasheet and real board testing. Prices and price movements reflect publicly visible channels — the quotation and contract govern (some figures remain TBC).
3. The decision line
Stay on RK3588 if you need
- 8K decode or output, HDMI 2.1, or more simultaneous displays than RK3576 supports.
- High-bandwidth expansion: PCIe 3.0 ×4 NVMe, 10GbE, multi-channel high-speed capture.
- Larger on-device models (7B-class int4) with a hard frame requirement.
- A tender that literally names RK3588 — then “better value” is irrelevant.
Move to RK3576 when you need
- Single 4K or several 1080p streams around 30 FPS.
- Fanless operation in an aluminium extrusion — you delete fans, ducts and noise remediation.
- A sub-thousand-yuan device BOM at a few hundred to a few thousand units.
- HMIs, retail boards and self-service terminals, where ports matter more than peak TOPS.
4. Stop comparing board prices: split the quote into four layers
| Layer | What it covers | Why it matters |
|---|---|---|
| Board | Hardware plus published BSP, volume-tiered | The only layer comparable across vendors — and the one most people mistake for total cost |
| Adaptation | Peripheral drivers, interfaces, images, validation reports | Decides whether your team spends three months closing gaps |
| Productisation | Carrier board, enclosure, thermal design, DFM, prototypes, certification support | The layer hobby-board ecosystems simply don’t offer |
| NPI | Pilot run, tooling, SOP, substitute parts, quality boundaries | Decides whether 20 prototypes can become 2,000 shipped units |
The habit worth building: judge the choice by project contribution margin, not board price. Money saved on silicon tends to resurface in mechanics, certification and field returns.
5. What that maps to in our line-up
- EDGE RK3576 AI Box — mid-tier vision inference and edge gateways, fanless first.
- RK3588 AI Box — 8K, multi-screen and heavy local inference.
- RV1126B SOM — low-power AI vision for IPC-class and battery cameras.
- RV1126B evaluation board — prove the model before locking the hardware.
Send us your compute, IO, power and target-cost list and we will map it onto the table above, including the boundary price and risk points for each layer. That is far more useful than a bare board quote.
FAQ
- Q：Can RK3576 be a drop-in replacement for RK3588?
- Usually not. The carrier board pin-outs and power design differ, so expect a board re-spin, and RKNN models need re-quantisation and validation. The real test is simpler: do you need 8K, high-bandwidth PCIe or tens of gigabytes of RAM? If not, an RK3576 design is typically a full cost tier below.
- Q：How much real inference performance do I give up?
- For YOLOv8s int8, public benchmarks put RK3588 around 41–46 FPS and RK3576 around 32–36 FPS, roughly a 20–25% gap at the same model and resolution. If your target sits near 30 FPS or below, RK3576 is more than enough — but validate with your own model.
- Q：How do we contain further chip price increases?
- Three levers: put a price-adjustment clause in the quote with a defined base date and validity window; agree substitute DDR parts early and qualify them; and trade volume forecasts or non-cancellable orders for tiered pricing. Negotiating board price alone does not solve it.
- Q：Which hidden costs should I budget for in a full device?
- At least five: mechanicals and thermal design (fanless or not), radio and safety certification such as SRRC / CE / FCC / RoHS, antenna and EMC remediation, NRE and adaptation engineering, and tooling plus pilot run. Together they often exceed whatever you saved on the silicon.