Author: Zihao
From October 12 to 16, the 2026 China International Industry Fair (hereinafter referred to as CIIF) was held at the National Exhibition and Convention Center (Shanghai). A landmark change at this year's CIIF is the inaugural establishment of a professional IC exhibition—the IC Ecosystem Show (ICES). Previous CIIFs focused on industrial equipment, automation, and manufacturing technologies. This year, a dedicated IC Ecosystem Show was specifically set up in Hall 5.2, gathering over 300 IC exhibitors covering the entire industry chain from chip design, wafer fabrication, packaging and testing, semiconductor equipment to core materials. On this comprehensive industrial platform, a direct dialogue window is built between the IC industry and factory application ends.
Walking among the various booths, it is easy to feel that large models and industrial intelligent agents have permeated every corner of industrial manufacturing. However, computing power, which is most critical in other application scenarios, is not the sole standard for measuring the applicability of chip solutions. Industrial sites have multiple hard constraints: response speed, long-term operational stability, data and system security, tolerance to extreme working conditions, and full-stack hardware and software iteration capabilities. Each is an unavoidable threshold. At this exhibition, major manufacturers showcased a wide array of skills, responding to these stringent industrial requirements with targeted solutions.
01. Diverse Practices of Edge Computing Power to Adapt to Complex Industrial Reality Demands
Industrial equipment generally has a service life of five to ten years, while large models and machine vision algorithms iterate rapidly. After hardware is finalized, how to keep up with AI capability updates without repeatedly revising the whole machine is a realistic proposition that almost all edge-side chip manufacturers need to address.
Axera Semiconductor made its appearance at CIIF with the theme "Boundless Smart Manufacturing, Computing Power Arrives". The exhibition area was divided into four zones: core computing power and technology foundation, industry solutions, software architecture and toolchain, and ecosystem and value co-creation. It concentrated on presenting full-scenario computing power solutions for physical AI, and officially launched the new AX8860 high-computing-power series on-site. It is introduced that this product is specifically built for the edge-side deployment of large models. It does not one-sidedly pursue peak paper parameters, but is designed around model deployment, on-site inference, and system collaboration. Adopting a 6nm process, it features 256 TOPS mixed-precision computing power, native support for FP4, memory bandwidth up to 273GB/s, and maximum support for 128GB LPDDR5X memory. The hardware is equipped with AXLink chip cascading and PCIe 5.0 high-speed interconnection, along with 64-channel video decoding and 16-channel video encoding capabilities, adapting to multi-channel visual perception and multimodal analysis tasks.
Qiu Xiaoxin, Founder and Chairman of Axera Semiconductor, stated: "We aim to be a long-term and stable computing power foundation for our customers, enabling enterprises to focus on business innovation without being distracted by computing power migration. Axera Semiconductor is committed to bringing intelligence to every site. The release of the AX8860 series high-computing-power chips is exactly the fulfillment of this commitment."
The exhibition area also displayed the DM0.5 general-purpose embodied brain independently developed by Yuanli Lingji based on Axera's high-computing-power platform. It can complete interface plug-and-pull quality inspection, general industrial insertion operations, and dexterous hand flexible cutting practical demonstrations, covering scenarios such as electronic assembly, precision manufacturing, industrial quality inspection, and flexible processing. Axera Semiconductor believes that for stringent industrial sites, the complete closed-loop of on-site intelligence must simultaneously meet four requirements: "seeing clearly, calculating accurately, reacting quickly, and executing stably." Only when on-site equipment achieves closed-loop processing of "perception-judgment-execution" can intelligence respond instantly at the forefront of events, truly transforming edge AI into a new quality productive force driving industrial upgrading.
CIX Technology brought a complete industrial implementation system covering the full chain of "chip-edge-system" to the exhibition. The company's independently developed high-performance agent SoC P1 serves as the core foundation. Adopting a 6nm process and Armv9.2 architecture, it integrates a 12-core CPU, a 30 TOPS dedicated NPU, and a 10-core GPU, forming a multi-core heterogeneous computing power of CPU, GPU, and NPU. For industrial scenarios, the SVE2 vector unit can accelerate large model inference. Features like TrustZone and Secure EL2 meet industrial control security requirements. 16-lane PCIe 4.0 provides industry-leading expansion capabilities, compatible with mainstream operating systems, providing a safe and reliable computing power base for industrial scenarios.
Zhou Jie, Chief Ecosystem Officer of CIX Technology, stated that the current period is a critical window for entering the industrial market. Traditional industrial solutions have weak AI capabilities, while Agentic AI and Physical AI bring new scenarios and new workflows, generating new demands for high-computing-power platforms. Meanwhile, related technologies are still in their early stages. CIX's platform has strong computing power and rich interfaces. After customers choose it, they can enjoy the long-cycle and long-tail dividends of the industrial market, which is conducive to startups entering the industrial track. Therefore, CIX Technology lists "industrial intelligence" as one of the core scenarios for its edge-cloud AI product solutions, deploying it alongside smart terminals, edge-cloud collaboration, and embodied intelligence.
CIX Technology displayed in the exhibition area the achievements co-built with industry chain partners in industrial OT control, intelligent vision, real-time control, quality inspection, and edge computing, with the industry chain synergy effect continuing to emerge. The company stated that it will take the AIoT industrial edge base and the scaled replication of quality inspection and intelligent vision as the foundation, and embodied intelligence as the growth engine, to continuously deepen co-construction with industry chain partners, making low-power, high-energy-efficiency edge intelligent computing power an important base for the implementation of new industrialization.
Rockchip set up seven exhibition zones around four major directions: industrial AI, robotics, industrial vision, and industrial control. It jointly displayed whole-set solution physical objects with ecological partners. On-site demonstrations included industrial AI inspection, production defect detection, SOP production line compliance detection, dexterous hands, and the RK1828 VLA edge-side robot motion strategy large model.
Yang Kai, Head of Rockchip AI Ecosystem, stated that industrial equipment has a long service life, but large models and machine vision algorithms iterate quickly, updating almost quarterly. If AI computing power is completely solidified within the main controller, subsequent upgrades require revising the whole machine, resulting in high costs and long cycles. Our "main control SoC + AI co-processor" dual-track architecture of computing and control well solves this pain point. The main controller is responsible for real-time control and peripheral management, while the co-processor is dedicated to edge-side model inference, facilitating quick upgrades for customers.
The RK182X series AI co-processor provides 20 TOPS computing power, equipped with high-bandwidth memory. It connects to the main controller via PCIe 2.0 and USB 3.0, fully supporting the local operation of various large models such as 3B, 7B to 27B LLMs and VLMs, and can expand computing power through multi-chip cascading.
Based on the dual-track architecture, industrial AI extends from image recognition to image-text understanding, process verification, and knowledge Q&A, covering scenarios such as inspection, quality inspection, SOP production line standard verification, and equipment operation and maintenance. Its advantages lie in multi-core parallel deployment of multiple inference tasks, supporting VLM to analyze complex on-site conditions and deploy inference to the production site, reducing data round-trips to the cloud, and providing expansion for adding AI capabilities to existing equipment in different production environments.
In terms of robotics, the booth presented three layers of capabilities: cerebellum motion control, brain VLA, and Agent closed-loop. Through multi-sensor data collaboration and perception-control coordination, it achieves a complete autonomous closed-loop from path planning, scene analysis to alarm handling.
D-Robotics participated in CIIF for the first time this time. The booth brought the Sunrise S100, S600 chips, and the RoboGo full-chain development platform.
According to Jiang Zhou, Vice President of Sales at D-Robotics, the S100 has already achieved batch implementation in many industrial projects. The S600's computing power can reach 560 TOPS, capable of smoothly running VLA embodied models like π0.5. The enterprise does not one-sidedly pursue paper TOPS parameters, but values the actual running performance after model migration more. RoboGo is a SaaS-form development platform, with modules supporting split selection, aiming to reduce the labor costs of industrial enterprises in building AI teams, and connecting the whole process from data collection, simulation, model quantization to edge-side deployment.
Focusing on industrial scenarios, D-Robotics sees three major industrial opportunities: First, the hardware acceleration standardization of industrial "integrated big and small brains", with domain controllers and AI Boxes maturing, reducing the board development and system integration costs for industrial whole-machine manufacturers, making it convenient to focus on production line scenario implementation. Second, the industrial data closed-loop spawns specialized division of labor, with segments in data collection, simulation, training and inference deployment, building pipelines based on unified standards to achieve industrial data staying in the factory and local secure iteration. Third, the enhancement of domestic industrial supply chain collaboration capabilities, with chips, sensors, and software toolchains fully supported, shifting industry competition to a contest of supply chain collaboration and engineering delivery capabilities.
At the same time, collaboration difficulties still exist in the industrial field: interface protocols between industrial sensors and computing power platforms are not unified, leading to many repeated adaptations; there is a lack of recognized industrial evaluation benchmarks, increasing the difficulty of selection and verification; industrial data interoperability infrastructure is weak, lacking formats and sharing mechanisms, restricting model iteration. Relying on the RoboGo industrial software base, D-Robotics connects the links of data, training, simulation, deployment, and feedback, promoting the industrial industry to move from point-to-point debugging to platform-based collaboration.
Dongfang Suanxin displayed the complete product matrix of the DF1000 software-defined 3D stacked near-memory computing high-computing-power chip at the booth, relying on architectural innovation to reduce dependence on advanced processes. It is understood that Dongfang Suanxin completed the training adaptation of mainstream large models in 7 days, with a 128-card cluster running stably, and has completed large-scale delivery, planning to promote thousand-card cluster delivery by the end of the year. From chips, accelerator cards, servers, super nodes to clusters, a complete chain is formed, completing the adaptation of mainstream large models such as Llama, StepFun, and MOE, completing technical verification at the engineering level.
Digital China joined hands with Intel to appear at the IC Ecosystem Show with the theme "AI for Smart Manufacturing, Chip Chain for Industry". Facing the whole industry of industrial manufacturing, it concentrated on displaying full-stack solutions covering industrial computing power infrastructure and industry intelligent applications. Facing high-load tasks such as machine vision, data analysis, AI training and inference, the Deshengda 4U server SERVER05 exhibited in the area is designed based on the Intel® Birch Stream platform, supporting dual sixth-generation Intel® Xeon® Scalable processors, providing rich PCIe 5.0 expansion capabilities and multi-specification hot-swappable storage configurations.
At the same time, multiple industrial edge computing whole machines were concentratedly displayed, covering products such as Dingsheng Intelligent MI-WCL, Weisen Computer PTLHAA1 V1.1, Yanzhi Technology SA5420, and Chuangshi Xunlian AD03. In addition to computing power and multi-channel communication interfaces, the equipment supports wide voltage and wide temperature operation, adapting to on-site conditions such as industrial control, smart gateways, and smart retail. Aiming at the heat dissipation and deployment needs of high-computing-power equipment, the specially exhibited Chaoji ServMAX® EO20-X4 liquid cooling model can be equipped with dual fourth-generation or fifth-generation Intel® Xeon® Scalable processors and two liquid-cooled GPUs, providing high-performance computing power support for scenarios such as smart manufacturing, autonomous driving, smart logistics, and simulation computing.
The 358H embodied intelligence solution, as a major highlight of this exhibition area, adopts the Intel 18A advanced process and the brand-new Xe3 architecture. Positioned as a robot computing platform "integrating big and small brains in one chip", it achieves single-chip heterogeneous integration of CPU+GPU+NPU. While outputting strong dense computing power, it relies on the X86 architecture to achieve real-time control, supporting multimodal AI inference of vision, voice, and large models, and can complete model conversion, quantization, and heterogeneous acceleration. The whole machine achieves silent and efficient heat dissipation through an integrated air duct, high static pressure fans, and a shock-absorbing and noise-reducing structure, which can be implemented in robot dogs, mobile mechanical arms, smart inspection, and various embodied robot scenarios.
Exhibits in terms of storage are equally rich. Solidigm's multiple high-speed SSDs cover PCIe 4.0 and PCIe 5.0 specifications, improving throughput performance for servers, data centers, and AI businesses. At the same time, wide-temperature solid-state storage products adapted to harsh conditions are also displayed, such as DapuStor AdgeStor 5 series M.2 SSD, widely applicable to core storage scenarios in industries such as data centers, automotive and transportation, smart manufacturing, industrial automation, and outdoor large-scale monitoring systems. In addition, Biwin CXL memory modules, DDR5 memory, and other products will also be displayed together, perfecting the storage-memory overall solution from industrial edge nodes to backend data centers, providing a solid data base support for enterprise key businesses, cloud infrastructure, and AI businesses.
Allwinner Technology concentrated on displaying a rich product matrix and collaborative ecosystem system for six core industrial scenarios, completely presenting the industrial ecosystem comprehensive capabilities of underlying chips, basic software, key hardware, security systems, and industrial collaboration.
Among them, focusing on the high-reliability and high-real-time application needs of industrial control equipment, multiple industrial PLC, HMI, and gateway products equipped with its T153 were concentratedly displayed, and a dynamic scheme for high-precision AD acquisition of electric power was demonstrated on-site, presenting the scaled implementation capability of T153 in mainstream industrial control scenarios such as logic control, motion control, and power acquisition. T153 integrates a quad-core Arm Cortex-A7 and a single-core RISC-V XuanTie E907 processor, supports DDR3 / DDR3L / DDR4 memory, and balances computing performance and real-time response. It is equipped with 3 Gigabit Ethernet interfaces, 2 CAN-FD and LocalBus, and integrates ISP. In addition, T153 provides rich interfaces such as 24-channel GPADC, 6 TWI, 30 PWM, and 10 UART, supports multi-core AMP architecture, and meets the high real-time needs in small and medium-sized PLC scenarios.
Allwinner Technology stated that at this exhibition, 14+ ecological partners participated in the same exhibition, forming a collaborative display of the hardware and software industry chain. Hardware ecological partners concentrated on presenting development boards and whole machine solutions. The development board display cabinet displayed over 50 industrial boards, covering Allwinner's mainstream industrial-grade chip platforms and supporting development solutions. Software ecological partners focused on operating systems, middleware, and industry applications. Multiple joint solutions ran dynamically, jointly providing mature system adaptation and application support for the chip platform. The hardware and software collaboration intuitively verified the implementation capability of Allwinner chips in industrial control scenarios and can support application development from evaluation and verification to mass production implementation.
02. Storage Hardware Iteration to Respond to New Workloads Brought by Industrial AI
After the popularization of edge-side large models, model weight loading, and KV Cache operations, and massive small file reading and writing bring new pressure to industrial storage. At this exhibition, three different technical routes can be seen, corresponding to the differentiated needs of industrial control ends, AI computing power, and high-reliability special conditions.
Biwin Storage concentrated on displaying six major storage solutions for AI + Industry, data communications, electric power, rail transit, smart security, and smart vehicles this time. The product matrix covers multiple forms such as SSDs, embedded storage, memory sticks, and storage cards, providing multiple temperature grade choices including industrial standard grade (-20℃ to 75°C), industrial wide-temperature grade (-40℃ to 85℃), and automotive grade (-40℃ to 105℃).
Key exhibits at this booth include the TGS601 AI monitoring ultra-large capacity SSD, the TGQ200 series PCIe Gen4 wide-temperature SSD with high-speed reading and writing, the localized power eMMC TGE408 equipped with the independently developed main control SP1800, and the newly released TDE408-R during the exhibition, which adopts the industry's first integrated architecture. Among them, the TDE408-R can flexibly configure system and business data partitions to build a more efficient industrial storage architecture.
Relying on the full-stack capabilities of independently developed main control, firmware, and integrated packaging and testing, Biwin Storage has made deep optimizations for industrial site conditions such as wide temperature, vibration, and abnormal power failure. It possesses technologies such as multi-level temperature control, complete data verification links, and power-down protection, balancing lifespan, data security, and long-term supply capabilities. Multiple products have achieved scaled implementation: the security monitoring series products adapt to multi-channel video concurrent writing and have been widely used in vehicle monitoring and park monitoring systems; TGE408 meets the national standard requirements of the power industry and has achieved KK-level shipments; the automotive series products have entered multiple whole vehicle manufacturers and Tier 1s.
Facing the edge AI wave, Biwin Storage is no longer limited to standardized device output, but turns to scenario-based storage solutions. Aiming at the high-frequency reading and writing pressure brought by edge-side large models, through firmware tuning, it optimizes wear leveling and garbage collection strategies, reduces performance fluctuations, and cooperates with edge-side AI chips to alleviate the computing power waiting problem. In the future, it will adhere to the route of platformization plus scenario customization, ensuring compatibility with old platforms while iterating products, helping industrial and automotive customers achieve localized implementation.
DapuStor launched a dual-mode enterprise SSD, integrating a high-performance pSLC partition inside a single SSD disk body, realizing a single hard drive with both a high-performance metadata partition and a large-capacity data storage partition. This architecture effectively solves the pain points of traditional solutions such as occupying many slots, solidified performance bottlenecks, and complex expansion ratios. Under high-concurrency business scenarios, the file metadata operation performance achieves several times improvement. This product is very suitable for typical scenarios such as AI inference KV Cache hierarchical storage, industrial massive small file reading and writing, and limited space in edge cabinets. It balances high performance and large capacity within limited hardware space, simplifies the hardware architecture of edge nodes, and reduces operation and maintenance complexity.
HikStor Technology focuses on the application of new storage MRAM in the industry. Compared with traditional Flash, MRAM has a 35-nanosecond level writing speed, an ultra-high 10¹⁴ times erase/write lifespan, a wide temperature range of -40℃ to 150℃, and radiation resistance characteristics. Superimposing these four advantages, MRAM is no longer "another kind of memory", but a technical solution that can simultaneously meet the requirements of speed, lifespan, temperature, and high reliability in critical data storage scenarios.
Therefore, the evaluation criteria for industrial storage are changing. It no longer only looks at nominal reading and writing parameters. Actual performance under high load, hardware and software collaborative adaptation, and the total cost of ownership over the equipment's full life cycle are taking on higher weight.
03. Building System Bases to Bridge the Intermediate Links from Chips to Production Lines
Chip computing power modules are only hardware components. To truly serve factory businesses, it is inseparable from operating systems, industrial buses, and control software to complete system transformation.
Kyland Technology relies on its independently developed Intewell real-time operating system, MaVIEW industrial control software, and AUTBUS industrial bus to build a complete localized software-defined control technology stack, achieving localization for the chamber control system of semiconductor machines.
Yu Hongping, Sales Director of East China Region, Semiconductor Business Unit of Kyland Technology, stated that this system achieves isolated coexistence of non-real-time Windows/Linux environments and hard real-time PLC environments through virtualization technology. Even if the Windows/Linux system crashes or gets a blue screen, it will not interfere with the real-time control link, and at the same time completes high-speed data interaction through shared memory. Through simplified architecture design, it achieves the integrated integration of IPC+PLC and system memory sharing, ultimately reducing the volume of the whole equipment by more than 30%, and synchronously reducing energy consumption and hardware costs by more than 20%. Combined with high-frequency data collection and industrial edge AI algorithms, this control system can converge the temperature control fluctuation of the semiconductor chamber to within 0.5 degrees Celsius. The AI fault prediction function can output warnings about 30 minutes in advance of actual anomalies, with a fault prediction accuracy rate reaching 90%.
Tsai Chi-nan, General Manager of the Industrial IoT Business Group of Advantech, believes that the first half of the AI industry focused on cloud large model training, while in the second half of industry value release, the core battlefield is at the edge. Industrial AIoT as a whole will go through five major development stages: "Internet of Everything, Data Fusion, Visual Insight, Assisted Decision-making, Collaboration and Full Autonomy", and finally move towards a new model of multi-AI Agent collaborative work. He believes that currently, consumer-end intelligent agent applications have been popularized, but there are still many realistic thresholds for implementation at industrial sites; the core foothold of industrial AI lies in scenarios, requiring facing customers' real business pain points, rather than simply pursuing technical concepts.
Combining a large number of implementation practice cases, Tsai Chi-nan listed the currently mature implementation scenarios: in the field of smart manufacturing, edge computing power supports production line defect visual inspection and dynamic optimization of process parameters, helping factories reduce the defect rate; the smart energy scenario relies on edge nodes to complete real-time scheduling of microgrids, ensuring stable energy supply; the smart transportation field relies on edge low-latency computing to achieve rapid decision-making of roadside perception.
Wang Zhiyu, Deputy General Manager of Sales at Alibaba DAMO Academy, pointed out the common pain points of the industry: after many enterprises obtain processor IPs, they face difficulties such as complex software adaptation and long product mass production cycles, resulting in the problem of "having IP, but difficult to implement mass production". In addition to computing power challenges, edge scenarios also encounter difficulties in storage main control - the PCIe interface iterates rapidly, and the main control not only needs to run through high bandwidth, but also must ensure deterministic latency under high concurrency; different manufacturers have a large number of customization demands for NAND scheduling and error correction strategies. The open instruction set architecture is of great value to the supply chain security and iteration rhythm of storage main control.
In response to the above pain points, XuanTie has created the FLEX customized platform and edge-side AI solutions, equipped with a complete three-layer software stack of RTOS, Linux, and Android. Through DSL description, RTL, toolchains, and verification environments can be automatically generated, encapsulating processor IP capabilities into a directly implementable productized platform. At the same time, facing SSD main control, it has launched multiple solutions, from consumer PCs to data center all-flash, adapting to diversified storage tasks, and widely serving four major product directions: SSD main control chips, large model edge gateways, programmable edge-side NPUs, and low-power industrial terminals, reducing the development threshold for industrial chips.
AMD launched the new generation Ryzen AI Max+ Pro 495 processor, which can locally run 320B parameter large models and support multi-agent parallelism. Ji Zhaohui, Vice President of Marketing for Greater China at AMD, proposed that the industrial dividends of the agent era do not belong to any single enterprise. The upstream and downstream of the industry chain must adhere to open collaboration, break the barriers between chips, software, and industry applications, transform general computing power capabilities into usable solutions for industrial scenarios, and jointly explore the huge market opportunities brought by edge intelligence.
This IC Ecosystem Show places the entire semiconductor industry chain under the scenario of new industrialization. Various hardware continues to iterate with supporting software and industry solutions, deeply penetrating quality inspection stations, robot control cabinets, semiconductor equipment, and factory servers, serving the intelligent upgrading of the manufacturing industry. And the realization of true value still requires the industry to continue to take root at industrial sites, completing polishing and verification in real business scenarios.