On August 5, 2026, Rockchip, a leading domestic waferless semiconductor design company, officially launched the RK3688, a new generation flagship AIoT application processor. Based on 8nm process technology, this chip integrates a high-performance CPU and a proprietary NPU, achieving 12 TOPS of computing power, and is specifically designed for scenarios such as edge large model inference, smart cockpits, robotics, and industrial vision. Industry analysis suggests that the release of the RK3688 will further accelerate the penetration of AI from the cloud to the edge, reshaping the existing smart hardware market landscape.
Performance Leap: A Computing Engine for Edge AI
The RK3688 utilizes TSMC's 8nm FinFET process. The CPU features an octa-core Arm Cortex-A78 architecture with a maximum frequency of 2.4GHz, representing an approximate 30% performance improvement over the previous generation RK3588. For the GPU, it integrates a Mali-G710 GPU, supporting 4K 120fps video encoding and decoding. The most significant feature is the built-in, self-developed NPU from Rockchip, which boosts computing power from 6 TOPS in the RK3588 to 12 TOPS. This doubles the integer precision computing power and natively supports Transformer models and mainstream deep learning frameworks.
According to Rockchip, when running a large language model with 7B parameters, the RK3688 achieves a generation speed of 15 tokens per second while keeping power consumption under 8W. This metric implies that edge devices no longer need to rely on the cloud to complete complex AI tasks such as voice assistant functions, document summarization, and image generation, significantly reducing latency and privacy risks.
Architectural Innovation: From SoC to System-Level Solutions
In addition to enhanced computing power, the RK3688 integrates rich interface resources, including PCIe 3.0, USB 3.2 Gen1, dual Gigabit Ethernet, and MIPI CSI/DSI, supporting multi-screen heterogeneous display and various sensor connections. Rockchip has simultaneously released supporting software SDKs and model deployment toolchains to help developers quickly deploy trained models to edge devices.
In terms of security, the RK3688 embeds an independent secure island processor, supporting hardware encryption/decryption, national cryptographic algorithms, and a Trusted Execution Environment, meeting the security compliance requirements for smart terminal devices in financial payment and smart park scenarios.
Market Context: AI Demand Explodes at the Edge, Domestic SoCs Seize Opportunities
As large model technology permeates various industries, AI inference is shifting from data centers to edge devices such as smartphones, automobiles, cameras, and robots. According to forecasts from market research firm Omdia, the global AIoT chip market size will exceed $12 billion in 2026, with edge AI inference chips accounting for 35% of that, up from 18% in 2024. However, this market has long been dominated by overseas manufacturers like Qualcomm and Nvidia, and domestic high-end SoCs have shown a clear gap in computing power and ecosystem.
Rockchip's consecutive launches of the RK3688 and the higher-tier RK3688S are aimed at filling this gap. Chen Feng, Executive Vice President of Rockchip, stated at the press conference, "The computing power of the RK3688 has reached mainstream international levels, and we offer extremely competitive cost-performance and localized services to help customers gain an advantage in smart hardware innovation."
Industry Interpretation: Intensifying Competition, Ecosystem Integration is Key Besides Rockchip, domestic competitors such as JCET and Allwinner are also accelerating the launch of new generations of AIoT chips. JCET's A series focuses on ultra-high definition and multimedia, while Allwinner focuses on low power consumption and vision processing. Compared to competitors, the RK3688 has a distinct advantage in NPU computing power. For example, JCET's latest chip, the A3212, has an NPU computing power of 8 TOPS, while Rockchip's 12 TOPS allows for running models with larger parameters.
However, market analysts point out that simply accumulating computing power does not guarantee success. Chip design companies need to provide a complete software stack and a developer community to lower the barrier to application development. Rockchip has a presence in the Linux, Android, and HarmonyOS ecosystems, and its open-source RKNN SDK has attracted over 50,000 developers, laying the foundation for building an ecosystem moat.
Application Prospects: From Smart Home to Humanoid Robots
The initial applications of the RK3688 will focus on smart cameras, smart control screens, AR/VR glasses, and industrial HMIs. In the intelligent cockpit sector, Rockchip has been jointly advancing in-cabin solutions based on the RK3688 with several automakers, implementing functions such as driver monitoring, gesture recognition, and voice interaction. Furthermore, in the humanoid robot field, the RK3688 can simultaneously handle visual perception, decision control, and motion control, making it a cost-effective "cerebellum" chip.
According to insiders, the RK3688 is scheduled for mass production in September 2026, and several leading ODMs have already completed design solutions, with terminal products equipped with this chip expected to hit the market in the first half of 2027.
Future Outlook: The Trend of High-End Domestic Chip Design
Leveraging over a decade of experience in multimedia processing SOCs, Rockchip is moving towards higher process nodes and stronger computing power. The release of the RK3688 is not just a product iteration but also a significant signal of Chinese chip design companies challenging the front-end AI battleground. As the domestic semiconductor supply chain gradually matures, we have reason to believe that key players, represented by Rockchip, will play a more important role in the global AIoT wave with "Made in China" chips.
OceanRing Asia Pacific Finance will continue to monitor subsequent market performance and test data.
