ChangXin Technology showcases storage products for three major scenarios in the AI era at IC China 2026
At IC China 2026, ChangXin Technology exhibited memory products including RDIMM, LPDDR6, and automotive-grade LPDDR5X for data centers, device-side AI, and smart vehicles. Company Vice President Luo Xiaodong said that storage competition has shifted from single-parameter comparisons to full-scenario product layouts, and noted that LPDDR6 has already entered mass production debut on the Xiaomi 18 Fold, the G5 platform has entered mass production, and G6 has completed engineering sample validation. In the first three quarters, ChangXin invested 11.071 billion yuan in R&D, up 73.03% year-on-year.
Why read this
AI score 52
At IC China 2026, ChangXin showcased three product lines—server, device-side, and automotive-grade—and provided progress updates on G5 mass production and G6 engineering samples, offering insight into the tiered layout approach of domestic storage.
Topics
Body · AI translation
On November 12, the 23rd China International Semiconductor Expo (IC China 2026) opened in Beijing. Memory manufacturer ChangXin Technology showcased a concentrated lineup of memory products for three core scenarios of the AI era—data centers, edge AI, and intelligent vehicles—including RDIMM, LPDDR6, and automotive-grade LPDDR5X.
That day, at the IC China 2026 forum, ChangXin Technology Vice President Luo Xiaodong pointed out that the competitive logic of the memory industry has undergone a fundamental shift, moving from the single-dimensional comparison of specifications in the past to competition in full-scenario product layout capabilities. Based on a deep understanding of application systems, creating differentiated memory solutions suited to different tiers and different scenarios is becoming the core variable reshaping the competitive landscape of the memory market, and it is also opening a new window for industry players to change lanes and overtake.
Image/ChangXin Technology Vice President Luo Xiaodong
"Affected by the surge in global computing power demand, the growth logic of each market has clearly diverged," Luo Xiaodong said in his keynote speech, "Exploring New Boundaries of DRAM Products in the Era of Computing Power Stratification." Servers give priority to meeting high-value needs such as AI large models and intelligent computing centers, directly driving sustained growth in demand for high-end memory and high-performance chips; the automotive market is driven by intelligence, and demand for automotive memory and automotive-grade chips is expected to grow at a compound annual rate of about 30% through 2030; AI features are penetrating faster in consumer electronics markets such as phones and PCs, but these markets still maintain pronounced cyclical fluctuations in consumption.
In terms of application stratification, different stages of AI large models have different memory requirements. Decode inference generates answers token by token, and the bottleneck is memory bandwidth; Prefill reads in and understands the input all at once, and the bottleneck shifts to computing power; large-parameter multi-card training is constrained by memory capacity and interconnect capabilities. No single memory product can be optimal in all directions at the same time, so the supply side is bound to move toward product differentiation.
In ChangXin's product matrix, server RDIMM was called by Luo Xiaodong the "first channel for computing power data transmission." More precisely, AI inference is reshaping the value of server memory, and RDIMM is upgrading from a "supporting component" of system main memory to the "core buffer" for long contexts of large models.
It should be noted that in the past, conversational AI could clear or transfer KV Cache once a round ended; Agentic AI, by contrast, involves high-frequency, long-horizon autoregression. After a user gives a task, the Agent autonomously unfolds a chain of thought, calls search, writes and runs code, self-corrects after errors, and continues calling tools. A single task may involve dozens or even hundreds of loops. A March 2026 Gartner study shows that Agentic workflows consume 5 to 30 times the tokens per task of a standard chatbot, and a single user request may trigger 10 to 20 sequential model calls. Memory usage continues to grow as the task runs and is no longer "released when done."
This clearly changes server configuration logic. In the past, training was dominant, GPU computing power was highly saturated, and the CPU-to-GPU ratio was often 1:4 or even 1:8; in Agentic AI inference and end-to-end business flows, to handle complex scheduling, the industry is increasing the CPU ratio, moving toward 1:2 or even 1:1. As system main memory directly connected to the CPU, the number of RDIMM slots per machine and the installation rate of high-capacity modules have multiplied accordingly.
Long-context inference highlights this trend even more. A single person's short conversation KV Cache is only tens of megabytes; when the context is extended to 128K or even 1M tokens, or when tens of millions of people are concurrent, the KV Cache of one active request can reach several GB to more than ten GB, and dozens of concurrent long sessions can overwhelm the GPU HBM of an entire server. Industry adopts tiered offloading: HBM retains only the very small amount of active context for the tokens currently being generated; data that the user has not accessed for a long time (cold KV Cache) is "offloaded" to hundreds of GB or even more of RDIMM main memory, so that it can be quickly pulled back when the session is reactivated; when it is that user's turn, the CPU issues instructions so that context data is moved from RDIMM to HBM for computation. RDIMM thus becomes the "core buffer" for long contexts of large models.
At present, ChangXin's RDIMM product line is advancing in parallel: DDR5 dies have completed the iteration from 24Gb to 32Gb; RDIMM modules achieve full coverage of mainstream capacities from 32GB to 128GB and speed grades from 5600Mbps to 8000Mbps; MRDIMM covers 128GB and 192GB, with speeds up to 12800Mbps, precisely addressing servers' and data centers' demand for high capacity and high speed.
Global mainstream server memory has entered a stable capacity-upgrade cycle, the share of low capacities is narrowing, penetration of mid- and high-capacity products is rising, and the center of value in the RDIMM market continues to shift toward high-capacity segments. Behind this trend is precisely AI inference pushing memory from a "capacity configuration item" to a "performance-critical item."
Image/ChangXin booth
Edge/automotive-grade capabilities stand out, scenario adaptation shows expertise
Edge AI is another main line of ChangXin's exhibition this time. In August this year, ChangXin officially launched LPDDR6, the industry's latest-generation LPDDR product, and it has already achieved global first mass production on the domestic flagship phone Xiaomi 18 Fold.
The product has a 16Gb capacity and a maximum speed of 12800Mbps, delivering a 60% increase in bandwidth and a 20% reduction in power consumption compared with the previous-generation LPDDR5X. Luo Xiaodong said that with its higher native speed and dual-die architecture design, ChangXin's LPDDR6 significantly improves the key metric of "bandwidth-to-capacity ratio," a feature that will enable it to expand in the future into intelligent driving, edge computing, and server scenarios with high requirements for memory energy efficiency.
On the LPDDR5X product line, ChangXin launched two new LPDDR5X products in 2026 based on its fifth-generation technology platform (G5), each with a single-die capacity of 24Gb, a 50% increase over same-type products of the previous generation. They use 496Ball and 245Ball package specifications, respectively, and can adapt to the differing scenario requirements of mid-to-high-end smartphones, portable consumer electronics, and more; both have now entered volume production. ChangXin LPDDR5X has already been deployed in the world's first AI agent phone, the Nubia NaviX Ultra, empowering an on-device AI interaction experience that "understands what you say, gets things done, remembers, and is secure enough."
The highlights in automotive scenarios are even more concentrated on the technical threshold. As is well known, automotive-grade memory is widely recognized in the industry as a high-barrier track, with core differences from consumer-grade solutions. The automotive-grade LPDDR5X that ChangXin is exhibiting this time combines high-speed transmission with large capacity—using bandwidth to support the timely caching and retrieval of perception data, providing a data foundation for processor perception fusion and algorithm computation; and using ample running space to carry cockpit programs, interfaces, and working data.
To respond to the integration design needs of automotive platforms, ChangXin brought two packaging solutions: one uses a 315-ball package, optimizing the I/O layout to achieve low parasitic parameters, making signal transmission more stable with less interference while also making the package smaller, balancing signal integrity and power efficiency; the other uses a 441-ball package, configured with 4 independent channels with a total data bit width of 64bit. The multi-channel parallel architecture supports high-speed parallel transmission and can carry greater data interaction throughput, meeting the high-bandwidth requirements of automotive domain control. The two solutions complement each other in connection layout and interface configuration, allowing customers to flexibly select models according to platform characteristics.
In terms of reliability, ChangXin's automotive-grade LPDDR5X memory products have passed authoritative certifications such as AEC-Q100 Grade 2 and obtained ISO 26262 ASIL D product functional safety certification, and the related management systems cover ISO 9001, IATF 16949, and QC 080000. At present, ChangXin's automotive memory products have steadily evolved from LPDDR4X to LPDDR5X, and the company is accelerating the application innovation and expansion of LPDDR6 in automotive scenarios.
Edge devices and automotive-grade are not a simple extension, but a "touchstone" of tiered capabilities. To this day, competition is shifting from "whether there is a product" to "whether it can be adapted to the scenario." Power consumption, packaging, reliability, and platform differentiation... each one tests manufacturers' depth of understanding of real applications.
Architecture breakthrough relies on technical trump cards, R&D investment increases to meet scenario needs
Having addressed the proposition of "what the market needs," how does ChangXin answer it? The key lies on the technical side.
ChangXin Technology President Cao Kanyu once pointed out at a public event in October this year that since its founding, ChangXin has initiated 4F² research and patent layout, with more than 300 published patent families, leading manufacturers at home and abroad in number, and in 2021 published mainland China's first paper related to the 4F² architecture.
The core feature of the 4F² architecture is ChangXin's self-developed VSG (Vertical Surrounding Gate) vertical surrounding gate transistor. It adopts a vertically surrounding channel structure to achieve stronger gate control and solve the pain points of leakage and short-channel effects in DRAM memory cells; the vertical transistor helps area scaling, and compared with the mainstream 6F² architecture, memory density in the same area can increase by 30%–50%. The accompanying hybrid bonding technology has been independently developed since 2019, with the number of published patents entering the global top ten; in 2025, a paper was published at IEDM, completing for the first time experimental verification of wafer-level bonding of DRAM and CMOS on a 12-inch wafer platform.
The technology roadmap is advancing in parallel. ChangXin stated at its third-quarter earnings briefing that, on the basis of key breakthroughs in the 4F² vertical channel DRAM architecture, the sixth-generation process technology platform G6 has completed engineering sample verification, which means that while the G5 platform is in mass production, engineering verification of the next-generation platform has already been put in place. ChangXin has designated the first-generation product of the 4F² architecture as DDR5 RDIMM and expects to launch a new-generation DDR5 RDIMM integrating the above advanced technologies by the end of 2026. Server-grade RDIMMs must support 24/7 uninterrupted operation, withstand the voltage impact of live hot-swapping, complete hundreds of stringent compatibility tests with CPUs and motherboards from different manufacturers and generations, and ensure that signals do not attenuate or produce bit errors on longer module traces during high-speed transmission. The new architecture will also drive collaborative innovation across design, process, equipment, materials, and other links, promoting transformation of the domestic memory industry chain.
High growth must be supported by high R&D investment. In the third quarter of 2026, ChangXin Technology achieved operating revenue of approximately 118.714 billion yuan, a year-on-year increase of 613.18%; and achieved net profit attributable to the parent company of approximately 54.435 billion yuan. In the first three quarters, it achieved operating revenue of 269 billion yuan, a year-on-year increase of 738%; and net profit attributable to shareholders of the parent company of 132 billion yuan. In terms of R&D investment, cumulative R&D investment in the first three quarters was 11.071 billion yuan, a year-on-year increase of 73.03%.
It is worth mentioning that at this exhibition, the fifth-generation technology platform (G5) already in mass production that ChangXin is exhibiting relies on quadruple patterning technology to scale the half-pitch of the active area of the memory array down to 11.95 nanometers, comparable to the DRAM mass-production process at the world's top level, and the number of dies produced per wafer is increased by at least 50% compared with the previous-generation platform, greatly enhancing product competitiveness and economic benefits.
Behind the LPDDR6 debut, G5 mass production, 4F² progress, and G6 engineering sample verification is ChangXin's continued increase in R&D investment and industry chain collaboration. Without an R&D foundation, a tiered product layout is difficult to sustain; with technical trump cards, changes in demand can be transformed into incremental space.
Final thoughts
Jiwei observes that what ChangXin is showcasing at IC China 2026 is not just a product matrix, but a clearer competitive logic. When the demand curves of servers, automobiles, and mobile phones are no longer synchronized, and when Decode, Prefill, and multi-card training in AI inference impose different constraints on bandwidth, computing power, and capacity, memory manufacturers must shift from "one product conquering the world" to "full-scenario product layout." And 4F², VSG, hybrid bonding, and the G5/G6 platforms are ChangXin's underlying technical trump cards; together they will constitute the "decisive factor" of competition in the next stage.