A product launch event themed "Awakening at the Edge and Endpoint, Intelligent Computing for the Future" has put a chip company founded just three years ago in the spotlight. On August 13, 2026, Wuxi Chengheng Microelectronics Co., Ltd. (Chengheng Micro) officially launched the CH37 series AI SoC, the company's first flagship product for edge and endpoint applications.
Edge and endpoint AI is an exploding trillion-yuan market. According to Frost & Sullivan, the global endpoint AI market size will leap from 321.9 billion RMB in 2025 to 1.2 trillion RMB in 2029, with a compound annual growth rate of up to 40%. In 2026, edge and endpoint native intelligence based on dedicated AI SoCs is reaching a critical turning point from technological exploration to mainstream industrial applications.
From Cloud to Edge and Endpoint: An Irreversible Migration of Computing Power
Over the past few years, AI development has been almost entirely concentrated in the cloud, with large-scale training and inference competing in the size of GPU clusters. However, AI computing is moving from a single cloud to a collaborative architecture integrating "cloud, edge, and endpoint". The cloud handles large-scale training tasks, while the edge and endpoint are responsible for low-latency inference, delivering intelligence directly to the physical world.
The driving force behind this shift comes from the industry side, with industrial advancement far exceeding expectations. Taking humanoid robots as an example, data from MIR shows that in the first half of 2026, China's humanoid robot production reached nearly 25,000 units, a year-on-year increase of 312%; shipments reached nearly 20,000 units, a year-on-year increase of 314%. The demand for computing power from terminal devices such as robots, drones, and smart vision is upgrading from basic usability to superior user experience.
Chu Libin, General Manager and CEO of Chengheng Micro
Chu Libin, General Manager and CEO of Chengheng Micro, stated: "Since its inception, Chengheng Micro has always focused on endpoint intelligence, rejecting the involution trend, and focusing on building a domestic AI SoC platform with low power consumption, high performance, and wide adaptability. The CH37 series significantly enhances the local inference capability of lightweight large models, achieving an ultimate balance between computing power and power consumption, and comprehensively empowering scenarios such as smart manufacturing, industrial vision, and edge computing."
"Heterogeneous Computing, Unified Architecture": Rebuilding Edge and Endpoint Computing Power with a Single Chip
Traditional endpoint AI systems mostly adopt multi-chip stacking solutions, where data is frequently transferred between chips, facing pain points such as transmission bottlenecks, high latency, and power dissipation. The toolchains of various chip manufacturers are fragmented, requiring developers to adapt to multiple toolchains, resulting in long R&D cycles and high deployment costs.
The answer provided by Chengheng Micro is "heterogeneous computing, unified architecture".
The core concept of the CH37 series is to integrate the dispersed functional modules in traditional multi-chip solutions into a single chip. The CH37 series integrates multiple computing units such as CPU (responsible for general computing and system scheduling), GPU (responsible for rendering and graphics processing), and NPU (responsible for AI inference acceleration) into a unified SoC design. In this way, end-to-end response performance can be significantly improved in terms of speed, and power consumption can be reduced by up to 30%.
More importantly, through a unified memory architecture, computing units such as CPU, GPU, and NPU can share data, reducing the overhead of massive data transfer and resource scheduling costs.
Specifically, the product specifications of the CH37 series are as follows:
CPU: It adopts a 10-core design, including 4 large and 4 small main cores, as well as 2 R8 real-time control cores. The R8 core frequency reaches 800MHz, responsible for global task decomposition and resource scheduling.
NPU: The hexa-core NPU provides 48 TOPS INT8 / 24 TOPS FP16 AI computing power, custom-designed for neural network inference, supporting local inference of lightweight large models.
GPGPU: The quad-core GPGPU provides 1 TFLOPS FP32 computing power, compatible with the CUDA ecosystem, offering massive thread parallel processing capabilities.
GPU: The self-developed rendering GPU provides 358 GFLOPS computing power.
DSP and Accelerators: It integrates a dual-core DSP and a self-developed FFT hardware accelerator, excelling in high-speed Fourier transform and filtering operations.
Perception and Storage: It is equipped with dual independent ISPs (including thermal imaging ISP algorithms), supports LPDDR5 up to 6400Mbps and 32GB addressing, and provides PCIe 4.0 interfaces and 10 Gigabit Ethernet ports.
Lin Cangsong, Deputy General Manager and CTO of Chengheng Micro
Lin Cangsong, Deputy General Manager and CTO of Chengheng Micro, emphasized at the launch event: "The five major computing units achieve seamless collaboration through a unified on-chip interconnect, breaking the dilemma of computing power islands in traditional multi-chip solutions. The system can automatically and accurately allocate tasks to the most suitable computing units: the CPU handles global control, the NPU processes neural network inference, the GPGPU executes large-scale parallel computing, and the DSP processes real-time signals, so developers do not need to care about the underlying hardware details."
In terms of actual measured performance, the CH37 series demonstrates the robust capabilities required for edge and endpoint scenarios.
The CH37 series achieves industry-leading levels in both computing performance and AI algorithm performance, achieving multi-fold improvements compared to mainstream competitors; it has excellent 16-channel 1080P video concurrent processing capability, and the end-to-end processing latency for dual-channel 1080P and 4K is controlled at an extremely low level, fully meeting the strict real-time requirements of edge and endpoint scenarios; each module adopts independent power supplies, and through intelligent DVFS for precise power control, power consumption can be reduced to less than half of that of similar products, while significantly reducing the number of peripheral components, optimizing PCB design, and greatly reducing the power consumption of the overall system.
The single-chip solution achieves a breakthrough balance in three dimensions: performance, power consumption, and area.
Delivering Not Just Chips, But Development Capabilities
Chengheng Micro has built a three-tier software architecture for the CH37 series.
The underlying layer is the heterogeneous computing foundational environment, which uniformly manages different computing units through unified scheduling management and type-based scheduling. Developers do not need to care about the underlying hardware topology, and the system automatically allocates tasks to the optimal computing platform. The middle layer is a complete SDK and ModelZoo model library, which encapsulates all hardware capabilities, covering mainstream scenarios such as vision and speech, achieving an "out-of-the-box" experience; the upper layer is a full-stack solution for specific applications, providing deeply optimized solutions for specific business scenarios.
"We are not just delivering a chip, but an out-of-the-box development capability," said Lin Cangsong. For customers with special needs, Chengheng Micro can also provide customized development support.
This full-stack delivery model of "chip + SDK + solution" enables partners to significantly shorten the R&D cycle from chip adaptation to solution deployment, reducing development costs.
At the launch event, multiple ecosystem partners shared their application explorations of the CH37 series.
Yuan Ruilin, General Manager of Shenzhen Nantian Donghua Technology Co., Ltd., demonstrated the application scenarios of the CH37 series in the financial industry: from real-time customer flow monitoring, anomaly analysis, and OCR recognition at branches to embodied robot services at branches, endpoint AI is reshaping the entire chain of financial services.
Xiao Guoqiang, Sales Director of HPCTek Inc, launched three hardware products based on the CH37 series on site: industrial control motherboards, industry edge computing boxes, and Lobster Computing Boxes, and demonstrated cross-industry solutions such as community security, road monitoring, industrial scenarios, and commercial and residential scenarios.
At the launch event, Chengheng Micro officially signed strategic cooperation agreements with key customers and partners such as Inspur Electronic, HPCTek Inc, Zhejiang Liangchi, and Siptechx. All parties will carry out in-depth cooperation around the CH37 series in fields such as AI edge computing, smart perception, and embodied intelligence.
Conclusion
Lin Cangsong gave a vivid metaphor at the launch event: the cloud is like a base station, the endpoint is like a mobile phone, and tokens are like data traffic. At present, the layout of cloud infrastructure is becoming increasingly complete, and the track belonging to edge and endpoint intelligence has just begun. The "Edge and Endpoint Awakening" has arrived, and the "Intelligent Computing Future" is being staged.
As AI computing power continues to migrate from the cloud to the edge and endpoint, whether the CH37 series, which has taken a differentiated route relying on "heterogeneous computing, unified architecture", can make a breakthrough in the trillion-yuan endpoint AI market is worthy of continuous attention.