Rockchip RK3572 / RK3576 / RK3588: In-Depth Specification Comparison and AI Agent Deployment Analysis

Summary

On the 8nm process node, Rockchip has laid out AIoT chips in three tiers. RK3588, the flagship platform released in 2022, is positioned for high-end edge computing and multimedia processing. RK3576, introduced in 2024, takes the upper-midrange role and carries flagship performance down into a broader demand. RK3572 is the midrange AIoT platform released most recently, in 2026, for tablets, AI cameras, NAS, retail terminals, and similar scenes. Together they form a complete product matrix of flagship, upper midrange, and midrange.

This article compares process, CPU, NPU, GPU, memory, and video codec item by item, and focuses on the practical constraints and selection advice for deploying AI agents on the three chips.

1. Product-matrix positioning and how the generations relate

The RK3572 model name sits close to RK3576, but the two are not a simple performance scale of each other. CPU architecture, NPU compute, and the GPU design differ substantially, while RK3572 is the one that supports the newer LPDDR5/5X standard on the memory interface. The technical comparison below goes layer by layer.

2. Core specifications, item by item

2.1 Process and CPU architecture

All three chips use a big.LITTLE heterogeneous design, and the generation gap is concentrated in the choice of big core. RK3572 uses A73 big cores, a mid-generation step in the ARM Cortex-A7x series, with single-core performance between A72 and A76. The A72 big cores of RK3576 are behind the A76 of RK3588 in absolute performance, but A72 energy efficiency in embedded use is mature and stable. The A76 of RK3588 is the performance-oriented flagship big core in the current ARM Cortex-A series. With a 3MB L3 cache, it has a clear advantage in AI inference tasks that need the CPU in the schedule.

For AI agent deployment, the CPU-architecture difference matters more than the surface parameters. KV Cache management, tokenizer processing, and multi-turn context maintenance in large-model inference depend heavily on single-core CPU performance. A76 big cores schedule these scenes more efficiently than A72 and A73.

2.2 NPU compute and precision support

The 4TOPS NPU of RK3572 is the lowest of the three. That compute level sits in the range described as strong on-device AI compute, but not enough to support inference of a large-parameter language model. Large language model deployment is not in the official list of application scenes. NPU compute for RK3576 and RK3588 is 6TOPS in both cases, and the precision ranges are largely the same. Actual inference performance cannot be read from the TOPS number alone. In measurements of the YOLOv5s (640×640) object-detection model, RK3588 reaches 49fps and RK3576 reaches 45fps, a gap of about 8%. That gap comes mainly from the triple-core NPU architecture of RK3588 and from higher memory bandwidth, not from a difference in total compute.

2.3 GPU and the memory subsystem

The GPU designs differ just as clearly. The Mali-G610 MP4 of RK3588 is far ahead of the other two in graphics rendering and general compute (OpenCL). That has practical value for vision-agent scenes that need the GPU to assist AI preprocessing, such as image normalization and color-space conversion. The Mali-G310V2 of RK3572 is ARM's latest entry-level GPU architecture. Its energy efficiency is strong, and its absolute performance is limited.

Memory bandwidth is an AI-inference bottleneck that is easy to underestimate. Theoretical bandwidth is about 34.1 GB/s for LPDDR4x, while LPDDR5/5X can reach 51.2 GB/s, an increase of about 50%. On the memory interface, RK3572 is the most aggressive of the three. It supports LPDDR5/5X natively, which means that in a real deployment its memory-bandwidth bottleneck can be smaller than the LPDDR4 base design of RK3576. The NPU compute ceiling (4TOPS) is the bottleneck that arrives sooner. Some studies find that in low compute-density inference, memory bandwidth often becomes the dominant constraint. Under heavy memory pressure, RK3588 can still keep latency variation within 5%. That property matters for the real-time response of an agent.

2.4 Video codec and ISP

RK3588 supports 8K@60fps decode and 8K@30fps encode, and it is equipped with a dual 16M Pixel ISP, with HDR and 3DNR. RK3576 supports 8K@30fps decode. RK3572 also supports 8K@30fps decode and 4K@30fps encode, and it integrates a 12-megapixel ISP. For an AI agent whose front end is visual perception, such as a video-analysis agent, the multi-channel 8K processing of RK3588 and its stronger ISP provide the hardware base for multi-camera input and high-quality image preprocessing.

3. Hardware constraints of AI agent deployment

An AI agent (AI Agent) is fundamentally different from traditional one-shot AI inference. An agent usually needs multi-turn dialogue, tool calls, memory management, and task planning. That means the hardware load is not a one-time model inference. It is a continuous loop of inference, decision, and inference again. The fit of the three chips is analyzed below on three dimensions.

3.1 Large language model inference

Through the RKLLM software stack, Rockchip provides large language model deployment support for RK3576 and RK3588. RKLLM Runtime loads the quantized model converted by RKLLM-Toolkit and runs LLM inference through the NPU driver.

At the model-support level, the 6TOPS NPU of RK3576 and RK3588 can run models in the 2B to 7B parameter class. Official tests show that, for the mainstream 2B-parameter models used on device, both chips can generate more than 10 tokens per second, which meets the basic need of deploying a small model at the edge. On measured data, in the RK3588/RK3576 + RK1828 dual-chip architecture Rockchip showed at WAIC 2026, first-token latency for the Gemma4 model is as low as 169.93ms, and mainstream 2B-7B models can all run smoothly on device.

In a real deployment, smooth running of a 7B model has a hard requirement on memory capacity. Take DeepSeek-R1 7B: on the RK3576 platform it needs more than 8GB of memory, while the 1.5B size can run on 4GB. A vision-dialogue model such as Qwen2.5-VL 3B also needs more than 8GB of memory. That means the 4TOPS NPU of RK3572 faces a compute ceiling in LLM inference. Even with plenty of LPDDR5 memory, 4TOPS of INT8 compute is hard pressed to support a usable inference speed for a 7B-class model. The AI positioning of RK3572 leans toward computer-vision inference (object detection, image classification, OCR) rather than an agent driven by a language model.

The difference in quantization strategy is also worth attention. On RK3576, W4A16 (4-bit weights, 16-bit activations) is commonly used to compress model size and to use the mixed-precision capability of the NPU. RK3588 more often uses a W8A8 path, and the NPU's native support for 8-bit quantization is more complete. In a real deployment, the balance of inference speed and accuracy for W4A16 on RK3576 needs extra tuning, while the W8A8 path on RK3588 is more mature.

3.2 Multi-task parallelism and CPU load

A typical AI-agent workflow is: receive user input → LLM inference produces an action → call a tool/API → process the tool result → LLM inference again. In this flow the NPU handles LLM inference, but the CPU has to handle a large amount of non-NPU work: tool-call I/O, context assembly, JSON parsing, and HTTP communication.

The advantage of the A76 big cores on RK3588 is decisive at this layer. While an agent is running, CPU load is not continuously saturated. It is pulsed. I/O waits during tool calls alternate with light compute. The higher IPC and larger L3 cache of A76 make the processing delay of these pulsed tasks lower, which shortens the agent's overall response time. Official material also states explicitly that "A76 big cores are better suited to running large AI models." That judgment holds more strongly for agents than it does for plain LLM inference.

The A72 big cores of RK3576 have the same 2.2GHz clock, but the architecture-generation gap puts IPC about 15%-20% below A76. In everyday dialogue that gap may not be obvious. When an agent has to call tools often, such as querying a database, operating the file system, or sending network requests, the accumulated delay on the CPU side grows sharply.

The A73 big cores of RK3572 are comparable to A72 in single-core performance, or slightly ahead, but there are only 2 big cores, and the 4TOPS NPU limit makes it hard to carry LLM inference and vision tasks at the same time. If the agent is designed as vision perception + a rules engine + lightweight NLP, RK3572 can be the perception front end. If an end-to-end LLM-driven agent is required, RK3572 is not a fit.

3.3 Memory capacity and the agent working set

An agent's memory need is far above a one-shot model inference. Besides the memory occupied by the model weights themselves, space has to be reserved for the following: the KV Cache, which grows linearly with dialogue turns; data returned by tool calls, which may include JSON, HTML, image descriptions, and similar content; and the agent-framework runtime, such as the memory overhead of frameworks like LangChain and Nanobot.

Take multi-turn dialogue with a 7B model deployed on RK3576: 8GB of memory is the minimum. In a real agent scene, the KV Cache of a multi-turn dialogue keeps occupying memory, and structured data returned by tools also accumulates in the context, so the margin on 8GB is quite tight. An RK3588 platform is usually paired with 8GB or 16GB of memory, and some core boards support up to 32GB, which leaves a more ample space for the agent's working set.

In addition, the triple-core NPU architecture of RK3588 allows more flexible NPU resource assignment. In principle, different NPU cores can be assigned to LLM inference and to vision-encoding tasks, so they do not contend. In Rockchip's "Lobster Agent" design, RK3588/RK3576 acts as the agent runtime sandbox and decouples compute from a dedicated AI accelerator such as RK1828, which uses exactly this schedulability of a multi-core NPU.

4. Selection advice: match the chip to the agent type

Scene 1: A pure vision-perception agent (object detection, image classification, simple OCR)

Recommended: RK3572.

The 4TOPS NPU of RK3572, together with the high bandwidth of LPDDR5/5X, is enough to run YOLO-series detection models and lightweight classification networks smoothly. The energy-efficiency advantage of the 8nm process makes it suitable for compact fanless devices. "AI cameras, video analytics, and retail devices" in the official positioning are typical scenes for this kind of application. Where cost is sensitive and language interaction is not required, RK3572 is the best fit.

Scene 2: A multimodal dialogue agent (vision + language, deployed locally)

Recommended: RK3576 (8GB+ memory).

On 6TOPS NPU compute, RK3576 is essentially level with RK3588 (the YOLOv5s inference gap is only 8%), while the price is about 30%-50% of RK3588. An 8GB LPDDR5 memory version can run vision-dialogue models such as Qwen2.5-VL 3B, and it also supports language dialogue at the DeepSeek-R1 1.5B level. For an agent project with a limited budget that still needs multimodal capability, RK3576 offers the best performance-to-cost ratio. The constraints to note: inference speed for a 7B-class model is clearly lower than on RK3588, and tool-call delay on the CPU side is higher.

Scene 3: A complex task agent (many tool calls, long context, high concurrency)

Recommended: RK3588 (16GB memory).

The A76 big cores, triple-core NPU, Mali-G610 GPU, and larger memory capacity of RK3588 together make a platform for agent deployment with no weak spot. In scenes that need frequent tool calls, a long dialogue history, and vision and language tasks at the same time, the combined advantage of RK3588 is not something RK3576 can stand in for. Its 8K video processing also gives hardware headroom to video-analysis agents. The cost is significantly higher, and power control needs a more careful thermal design.

Scene 4: A lightweight agent aimed at energy efficiency

RK3572 can act as a perception coprocessor and form a heterogeneous architecture with RK3588/RK3576. RK3572 handles continuous visual perception, such as face detection and object tracking, and the main SoC handles LLM inference and task decisions. This split uses the low power of RK3572 and the inference capability of RK3588/RK3576. Similar practice already exists in Rockchip's "dual-chip architecture" designs.

5. Closing

The core difference among the three chips comes down to one judgment: NPU compute sets the agent's thinking speed, single-core CPU performance sets its responsiveness, and memory bandwidth and capacity set its working space.

RK3572 is the value choice for visual perception. A 4TOPS NPU and LPDDR5 support make it a strong front end for a vision agent, and it cannot carry an LLM-driven agent on its own. With a 6TOPS NPU and a highly competitive price, RK3576 covers most multimodal dialogue agents, and it is the balanced choice that is capable enough at a lower price. With A76 big cores, a triple-core NPU, and ample memory expansion, RK3588 is the only platform that can stably run a complex task agent on device. Its premium corresponds to higher agent response quality and a higher ceiling on task complexity.

Looking at the 2026 product cadence, Rockchip is decoupling the compute demand of agents through coprocessor designs such as RK3572 + RK1828. The main SoC handles scheduling and light inference, and a dedicated accelerator handles the heavy model. This trend means future agent-hardware selection will be an architecture-design question of "host + coprocessor," rather than a decision about a single chip.

The published specifications of AIBox03 are an octa-core up to 2.4GHz, Mali-G610, a 6T triple-core NPU, and 8GB LPDDR4, in the same class as the RK3588 tier in this article. The chip model follows the product page and the configuration as shipped.

If a 4 TOPS host does not run the large model by itself and hands inference to a coprocessor, see the RK3572 + RK1828 dual-chip architecture.