RK3572 is Rockchip’s mid-range 8nm AIoT processor for edge computing, personal mobile internet devices, and embedded applications that need multimedia and local AI performance. It combines an octa-core CPU, an NPU rated at up to 4 TOPS INT8 with sparsity, Mali-G310 graphics, hardware video processing, modern memory support, and extensive industrial I/O. For developers, its value lies in balancing processing capability, power efficiency, connectivity, and board-level flexibility.
What Is the RK3572 and Why Does It Matter?
The RK3572 is an 8nm Rockchip system-on-chip designed for edge AI, multimedia terminals, industrial equipment, and other AIoT products. It extends Rockchip’s mid-range platform family with an eight-core CPU configuration, a 4 TOPS NPU, and broad display, storage, and connectivity support.

Application illustration only; this is not a photograph of an RK3572 board or ArmSoM Zero5.
A mid-range processor built for edge AI
Rockchip describes the RK3572 as a low-power, high-performance processor for edge computing, personal mobile internet devices, and other AIoT applications. This positioning matters because many embedded products must process data locally rather than sending every camera frame, sensor reading, or user interaction to a remote server.
The integrated 4 TOPS NPU is intended for supported artificial intelligence workloads such as image classification, object detection, and other inference tasks. The datasheet qualifies the 4 TOPS INT8 figure with sparsity. Actual performance depends on the model, software toolchain, quantization method, memory bandwidth, and board implementation. TOPS is therefore a capability indicator, not a guaranteed application-level speed measurement.
The RK3572 is also relevant to multimedia systems. Its decoder supports H.265, VP9, AV1, and AVS2, with decoding capability listed up to 8K at 30 frames per second or 4K at 120 frames per second. H.264 decoding is listed up to 4K at 60 frames per second, giving the chip a foundation for digital signage, video terminals, and industrial displays. These codec-specific limits are documented in the Rockchip RK3572 datasheet, revision 1.1.
CPU, GPU, and platform balance
The CPU uses an octa-core configuration consisting of 2× Arm Cortex-A73 cores and 6× Arm Cortex-A53 cores. Published module specifications list the Cortex-A73 cores at up to 2.2 GHz and the Cortex-A53 cores at up to 2.1 GHz; operating frequencies depend on the selected board and software. This arrangement combines performance-oriented cores with efficiency-focused cores for operating-system tasks, application logic, networking, and concurrent background services.
Graphics are handled by the Mali-G310V2 MC1. The revision 1.1 datasheet lists OpenGL ES 1.1/2.0/3.2, OpenCL 3.0, and Vulkan 1.4. Its revision history explicitly updates Vulkan 1.2 to Vulkan 1.4, explaining why older sources may show 1.2. Developers should still confirm the graphics driver and SDK support on their selected board.
Who should consider it?
RK3572 is worth evaluating when a product needs several capabilities in one platform:
Local AI inference through the NPU.
High-resolution video decoding and display output.
Camera processing supported by a 12-megapixel ISP.
Industrial interfaces such as CAN, UART, I2C, SPI, and I3C.
Storage options including UFS, eMMC, SD, PCIe, or SATA.
Two gigabit Ethernet controllers for connected edge equipment.
Linux or Android development, with distribution support depending on the board.
Its feature set fits smart terminals, machine-vision equipment, digital signage, industrial edge devices, and other embedded AI products. It is a platform for systems running an operating system and application stack, rather than a substitute for a simple microcontroller in every design.
RK3572 Specifications at a Glance
The most useful way to evaluate RK3572 is to connect each specification with a design decision. CPU cores affect operating-system and application responsiveness, the NPU affects local inference, video blocks affect media pipelines, and I/O determines how the chip connects to sensors, displays, storage, and networks.

Board-design illustration only. The pictured board and measurement do not represent RK3572 or Zero5 hardware specifications.
Core hardware reference table
Category | RK3572 capability | Practical relevance |
Process technology | 8nm | Balances processing capability and power requirements for embedded products |
CPU | 8 cores: 2× Cortex-A73 + 6× Cortex-A53 | Operating systems, applications, networking, and multitasking |
CPU frequency | Published module figures: A73 up to 2.2 GHz; A53 up to 2.1 GHz | Confirm the operating points on the selected board |
AI accelerator | Up to 4 TOPS INT8 with sparsity | Supported on-device neural-network inference |
GPU | Arm Mali-G310V2 MC1 | Graphics, display composition, and supported GPU compute |
Graphics APIs | OpenGL ES 1.1/2.0/3.2, OpenCL 3.0, Vulkan 1.4 in datasheet revision 1.1 | Confirm matching driver and SDK support |
Video decoder | H.265, VP9, AV1, AVS2: up to 8K@30 or 4K@120; H.264: up to 4K@60 | High-resolution playback and multi-format video terminals |
Video encoder | H.264/H.265: up to 4K@40 in the Rockchip datasheet | Camera recording, streaming, and video processing |
ISP | 12-megapixel image signal processor | Camera and machine-vision pipelines |
Display interfaces | HDMI 2.1, eDP 1.3, MIPI DSI 1.2, parallel RGB/BT.656/BT.1120, E-ink | Monitors, embedded panels, and electronic-paper products |
Memory | LPDDR4/LPDDR4X-4266; LPDDR5/LPDDR5X-5500 | Memory bandwidth and power choices for board designers |
Memory addressing | Up to 16 GB; up to 4 ranks, subject to datasheet configuration restrictions | Larger applications, buffers, and concurrent services |
Storage | UFS 2.0, eMMC 5.1, SD 3.0 / MMC 4.51 | Embedded boot media and removable storage |
Expansion | PCIe 2.1 and SATA 3.1 | Compatible expansion devices and storage |
USB | USB 3.0/2.0 with dual-role device capability | Peripheral, service, host, and device connections |
Networking | Two 10/100/1000 Mbps Ethernet controllers | Dual-network edge-device designs; board-level PHYs and connectors required |
Embedded interfaces | SDIO 3.0, DSMC, CAN, I2C, I3C, SPI, UART | Sensors, controllers, and industrial peripherals |
This table describes SoC capabilities, not a promise that every RK3572 board exposes all interfaces. HDMI and eDP share an interface, and high-speed PCIe/SATA/USB resources also have multiplexing constraints. Check the datasheet, schematic, and board documentation together.
How to interpret the multimedia figures
Video specifications matter when a product combines capture, processing, encoding, and display. Decoder figures describe handling compressed input streams; encoder figures describe creating compressed video. They are separate hardware functions and must be evaluated against the intended pipeline.
For H.264/H.265 encoding, the Rockchip revision 1.1 datasheet lists 4K@40fps. The reviewed Forlinx RK3572 page contains both 4K@40fps in its description and 4K@60fps in its specification table. This guide uses the chip manufacturer’s 4K@40fps specification; it does not assume the conflicting value represents an additional supported mode. Verify the exact codec, driver, and hardware before committing to a 4K60 recording requirement.
Memory, storage, and connectivity choices
RK3572 supports LPDDR4, LPDDR4X, LPDDR5, and LPDDR5X memory. Its controller provides up to 16 GB of addressing space and up to four ranks, with configuration restrictions detailed in the datasheet. More memory can help when a system runs a graphical interface, local database, AI runtime, video services, and network applications together.
UFS 2.0 and eMMC 5.1 support onboard storage designs, while SD support can simplify development and field service. PCIe 2.1 and SATA 3.1 provide expansion options. Two Ethernet controllers are useful where one network connects to operational equipment and another connects to maintenance or data services. Which interfaces can operate together depends on the board layout and pin multiplexing.
How to Build a Product Around RK3572
A practical RK3572 project starts with workload definition, then validates the board, software stack, interfaces, and thermal behavior. Choosing a processor only from headline TOPS or video figures can cause problems if the required camera, display, driver, or storage path is unavailable on the chosen hardware.
Step 1: Define the actual workload
Write down the complete processing chain before selecting a board. A machine-vision terminal might need camera input, image preprocessing, neural-network inference, result display, Ethernet communication, event logging, and optional video recording.
Separate the workload into four groups:
CPU tasks: operating-system services, user interfaces, networking, scripting, and business logic.
NPU tasks: supported neural-network inference, with preprocessing and postprocessing evaluated separately.
GPU tasks: graphical interfaces, visualization, compositing, and supported compute operations.
Media tasks: camera capture, video decoding, encoding, display output, and storage.
This breakdown prevents a common design error: assuming that a 4 TOPS NPU accelerates every AI-related operation. Unsupported operators, data conversion, and preprocessing may still run on the CPU or GPU.
Step 2: Match the board to the interfaces
Compare the required connections with the selected RK3572 board or module. A display product may need HDMI, eDP, MIPI DSI, or an embedded-panel interface. A machine-vision product depends on camera connectivity and the ISP path. An industrial controller may prioritize CAN, UART, I2C, SPI, I3C, and Ethernet.
The processor’s interface list is broad, but a board cannot necessarily expose every interface simultaneously. Pin multiplexing, connectors, power circuitry, external components, and layout determine what is physically available. Treat the board schematic and carrier-board documentation as the final reference for a product design.
Step 3: Validate software and performance
Check software before finalizing hardware. As a vendor-specific example, Forlinx lists Linux 6.12 and Android 16 for its RK3572 platform, while Forlinx Desktop 24.04 and Debian 13 are marked as in development on the reviewed page. These are not ArmSoM software-support commitments. Other boards may have different kernels, bootloaders, graphics drivers, and AI runtimes.
A useful validation sequence is:
Boot the intended operating system from the target storage configuration.
Confirm display, USB, Ethernet, camera, and serial interfaces.
Convert and run the intended AI model using the target board’s supported NPU toolchain.
Test video decoding and encoding at the required resolution and frame rate.
Measure sustained performance with realistic simultaneous workloads.
Check standby, playback, and graphics power consumption on the complete device.
This process produces evidence that is more useful than a processor-level specification alone.
Practical Tips, Mistakes, and Real Examples
The RK3572’s broad feature set makes it flexible, but integration needs careful validation. Project risks include mismatched software support, unclear media requirements, insufficient cooling, or assuming that a chip-level specification applies unchanged to every board.
Common mistakes to avoid
Treating TOPS as an application benchmark. The 4 TOPS figure is an INT8 peak rating with sparsity. It does not guarantee camera frame rate, detection latency, or model compatibility. Test the exact model and input pipeline.
Assuming all video limits are identical. Decoding, H.264/H.265 encoding, and JPEG processing have different limits. Use the manufacturer’s 4K@40fps H.264/H.265 encoding specification as the reference, then validate the required path on the selected board.
Ignoring interface multiplexing. A board may expose only a subset of the processor’s interfaces. Confirm connector availability, shared resources, and pin assignments before designing a carrier board or enclosure.
Choosing memory without considering the software image. A graphical interface, containers, media buffers, and an AI runtime can consume substantially more memory than a minimal embedded Linux application. Validate the complete software image.
Overlooking sustained thermal behavior. A short benchmark can show peak performance while continuous video or AI processing produces a different result. Test the longest realistic workload in the intended enclosure.
Example: smart visual terminal
A smart terminal can combine a display, camera, local inference, network connectivity, and media playback. RK3572 offers relevant building blocks through its 12-megapixel ISP, NPU, Mali-G310 graphics, display interfaces, and hardware video decoder.
An illustrative design could use the ISP for camera processing, the NPU for a supported detection model, the CPU for application logic and networking, and the GPU for the user interface. If it also records video, confirm the encoder requirements on the selected board. This is an application example, not a report of a tested RK3572 product.
Example: industrial edge controller
An industrial edge controller may collect sensor data, communicate with machines, run local rules or inference, and forward selected results to another system. CAN, UART, I2C, I3C, SPI, Ethernet, PCIe, SATA, and USB provide relevant connection options.
For this use case, multimedia may be secondary. Software maintenance, I/O availability, boot reliability, storage, and enclosure behavior can matter more. A board exposing the required industrial interfaces can be more useful than one emphasizing display features the product will not use.
Which RK3572 Setup Fits Your Situation?
RK3572 can serve several product types, but the best board depends on workload, operating system, interfaces, and development stage. The same SoC may be used differently in a prototype, vision appliance, signage player, or industrial controller.
For developers building a prototype
A development board with accessible USB, Ethernet, display, storage, and serial connections is a practical starting point. Prioritize documentation, operating-system availability, camera support, and access to the NPU software stack.
Linux can suit services, automation, and hardware testing, while Android may suit touch terminals and consumer-style interfaces. Confirm the exact distribution and image status: a planned Debian release is different from an image available for download today.
For AI and machine-vision projects
Start with the model and camera requirements. Confirm supported operators in the RK3572 NPU toolchain, then test input resolution, preprocessing, inference latency, and postprocessing on the target board.
The ISP and NPU provide a starting point for local vision processing. Results also depend on camera connectivity, memory bandwidth, model conversion, lighting, image format, and whether recording or display composition runs at the same time.
For signage, media, and industrial products
Digital-signage projects should focus on decoder support, display output, storage, networking, and long-duration stability. Hardware decoding can reduce reliance on the CPU, but the exact playback application and codec profile still require testing.
Industrial projects should prioritize their required buses, Ethernet, storage, and software-maintenance path. ArmSoM has announced plans for ArmSoM Zero5, a compact RK3572 single-board computer concept targeting a Zero-style form factor. Read the ArmSoM Zero5 concept preview for the proposed layout. Final specifications, interfaces, software images, pricing, and availability remain subject to product validation and announcement.
Rockchip’s RK3572 announcement is dated 18 May 2026. It reports more than 100% higher performance and over 50% lower typical power than its previous mid-range platform, plus power figures below 10 mW in second-level standby, 670 mW during 1080p playback, and below 1 W under light GPU load. These are supplier measurements, not ArmSoM benchmarks. Rockchip specifies a tablet test configuration with 8 GB LPDDR5 and 128 GB UFS, the display and touch panel removed, and Wi-Fi/Bluetooth disabled. They should not be treated as whole-board power promises for an SBC with active peripherals. See the original Rockchip announcement and test conditions.
FAQ
What is the RK3572 used for?
RK3572 is designed for edge computing and AIoT products, including smart terminals, machine vision, digital signage, multimedia systems, and industrial edge equipment. Its CPU, NPU, GPU, media engines, ISP, and interfaces can support application logic, local inference, imaging, and network communication. The best fit depends on a board’s exposed interfaces and software support.
How powerful is the RK3572?
RK3572 combines two Cortex-A73 cores and six Cortex-A53 cores; published module specifications list the A73 cores at up to 2.2 GHz. Its NPU is rated at up to 4 TOPS INT8 with sparsity, and it includes Mali-G310 graphics. The datasheet lists H.265/VP9/AV1/AVS2 decoding up to 8K@30 or 4K@120, with H.264 up to 4K@60. Real performance depends on software, memory, cooling, and workload design.
Does RK3572 support Linux and Android?
Yes, RK3572 board vendors offer Linux and Android platforms. For example, Forlinx lists Linux 6.12 and Android 16, while its Desktop 24.04 and Debian 13 options are marked as in development. Verify the actual board’s downloadable images, kernel, graphics drivers, camera support, NPU runtime, and maintenance plans. One vendor’s software support does not establish support for every RK3572 board.
Is RK3572 suitable for machine vision?
Its 12-megapixel ISP, NPU, CPU, video engines, and interfaces make RK3572 a candidate for machine-vision designs. Suitability depends on the camera connection, supported AI model, resolution, frame rate, preprocessing, lighting, and sustained thermal performance. Benchmark the complete capture-to-result pipeline rather than relying on the TOPS rating alone.
Sources and Further Reading
Rockchip RK3572 datasheet, revision 1.1 — manufacturer document hosted on rockchip.fr.
Forlinx FET3572-C / OK3572-C product information — vendor-specific software and frequency example.
For an RK3572-based project or custom embedded design, contact ArmSoM with your required interfaces, operating system, AI model, and deployment constraints.
