Chinese memory manufacturer CXMT is reportedly exploring customised 3D DRAM and advanced 3D integrated circuit designs as a possible alternative to conventional high-bandwidth memory.
The strategy could help the company develop memory products for AI systems despite US restrictions limiting its access to some of the tools needed for advanced HBM production.
Rather than trying to copy the same path taken by Samsung, SK hynix and Micron, CXMT may focus on specialised memory built for specific processors and workloads. This approach would involve closer integration between memory and logic, potentially improving bandwidth, latency and power efficiency.
The opportunity is still uncertain. CXMT has reportedly produced sample 3D DRAM chips and is working on customer projects, but commercial yields, production costs and real AI performance remain unknown.
| Area | Reported position |
|---|---|
| Main strategy | Customised 3D DRAM and 3D integrated circuits |
| Reason for shift | Restricted access to advanced HBM manufacturing equipment |
| Target market | AI inference and specialised computing |
| Main competitors | Samsung, SK hynix and Micron |
| Current stage | Samples and active customer projects |
| Key risks | Cost, yields, performance and limited demand |
| Potential advantage | Faster entry into a specialised market |
US restrictions make conventional HBM development more difficult
HBM has become one of the most valuable components in modern AI accelerators.
The technology stacks several memory dies vertically and places them close to the main processor, providing much higher bandwidth than standard DDR or LPDDR memory. This allows large AI models to move data quickly between memory and compute units.
Producing competitive HBM requires advanced manufacturing equipment, packaging technology and process control.
US export restrictions reportedly limit CXMT’s access to some of the tools needed to develop and manufacture leading HBM products. This makes it difficult for the company to compete directly with established suppliers using the same design and production methods.
Customised DRAM could offer a different route.
Instead of manufacturing a standard HBM product for a broad market, CXMT could work with chip designers to create memory specifically matched to their processors, packaging systems and workloads.
This may allow the company to use technologies and equipment that remain accessible while still addressing the growing demand for faster AI memory.
3D DRAM could place memory and control logic on separate wafers
The reported strategy involves 3D DRAM and 3D integrated circuits.
In a conventional memory chip, storage cells and control logic are typically manufactured together within the same design. A 3D approach can separate these functions onto different wafers or dies before connecting them vertically.
The memory cells can be produced using a process optimised for density, while the control logic can use a different process designed for speed and power efficiency.
Advanced bonding can then connect the layers through short, dense electrical paths.
This could reduce the distance data must travel and allow more memory capacity to fit within a smaller package.
It may also improve power efficiency because less energy is needed to move data between separate components.
For AI inference, where models repeatedly read large quantities of weights from memory, these improvements could be valuable.
However, the technology must still prove that it can match or exceed HBM in practical workloads.
Samsung and SK hynix are taking a more cautious approach
Reports suggest that some fabless semiconductor companies have asked Samsung about customised 3D memory designs.
Samsung and SK hynix are believed to be researching similar technologies, but both companies appear cautious about rapid commercialisation.
Their current business models are built around high-volume production of standard DRAM and HBM products.
A customised product may require different designs for each customer, smaller production runs and closer technical cooperation. This can increase development costs and reduce the economic benefits of manufacturing at very large scale.
It could also compete with existing HBM products that already generate strong revenue and margins.
CXMT faces fewer concerns about protecting an established HBM business because it has less access to that market. This may give the company more reason to accept the risks associated with specialised memory.
The slower approach taken by Korean manufacturers could therefore create an opening for CXMT if customers are willing to adopt less conventional designs.
Custom memory could suit China’s domestic AI processors
China’s semiconductor industry includes a growing number of companies developing AI accelerators, server processors and specialised inference chips.
These products may not have reliable access to the same HBM supply used by major US chip companies.
Customised memory could be designed around the exact bandwidth, capacity, packaging and power requirements of domestic processors.
This would allow chip and memory designers to work together instead of selecting from a limited range of standard components.
Such cooperation may also support systems that combine logic and memory more closely.
For inference workloads, the design could focus on keeping model weights near the processing units and reducing the energy spent moving data.
The approach may be particularly useful for processors designed for a narrow set of models or applications.
It would be less attractive for customers seeking interchangeable components that can be sourced from several suppliers.
Production yields and costs remain major risks
Building sample chips is different from manufacturing them economically at scale.
Vertical bonding and multi-layer designs introduce more production stages. A defect in one layer or bond can affect the entire completed package.
Low yields could make each usable chip extremely expensive.
Custom products also spread engineering costs across fewer units than standard DRAM. This can make them difficult to price competitively unless customers commit to large orders.
CXMT must prove that its designs offer enough performance or supply security to justify those costs.
The company also needs to demonstrate reliability, thermal stability and consistent production quality.
AI systems often operate continuously under heavy load, so memory failures can be costly for data centre operators.
The available information does not confirm expected production dates, target customers or final specifications.
3D DRAM may not replace HBM across every workload
HBM benefits from a mature ecosystem and clearly defined interfaces.
Accelerator companies already design products around its bandwidth, packaging and software behaviour. Switching to a customised memory architecture may require major changes to the processor and system design.
This means 3D DRAM is unlikely to replace HBM universally in the near term.
It may instead serve selected customers that cannot obtain enough HBM or that need a memory system tailored to a specific accelerator.
The technology could also complement HBM rather than replace it. Some systems may use customised memory for capacity or inference while retaining HBM for workloads requiring maximum bandwidth.
The commercial opportunity therefore depends on how many customers value customisation over standardisation.
CXMT is taking a higher-risk route with potential strategic value
CXMT’s reported plan reflects the constraints facing China’s semiconductor industry.
Waiting for restrictions to ease would leave the company further behind established HBM suppliers. Developing a different memory architecture gives it a chance to compete in an area where the market is less mature.
The strategy also aligns with China’s broader effort to build domestic alternatives across processors, memory, manufacturing and packaging.
Success is far from guaranteed.
CXMT must overcome technical, manufacturing and commercial risks while proving that customised 3D DRAM can deliver useful AI performance.
Samsung and SK hynix could also enter the market quickly if demand becomes clear, using their larger manufacturing scale and deeper memory expertise.
For now, CXMT appears willing to move earlier and accept more uncertainty. That could give it a temporary advantage if specialised AI memory becomes a meaningful market, but the result will depend on production yields, customer adoption and whether the technology can compete with HBM in real systems.
Discussion (0)
Be the first to comment.