top of page

Samsung’s CXL Strategy Why Composable Memory Matters

23 hours ago
8 min read

Memory has become one of the hardest limits in modern computing. Processors keep getting faster, artificial intelligence models keep getting larger, and databases keep more information close at hand. Yet a server still often treats memory as something fixed inside one machine. If one server has too much and another has too little, that mismatch wastes money and power.


Compute Express Link, usually shortened to CXL, is one of the main industry answers to that problem. It gives servers a standard way to connect processors, memory, and other devices so memory can be added, shared, and managed more flexibly.


That is why Samsung’s CXL work matters. The company is not only selling memory chips. It is trying to shape how future data centers add and use memory.


Wide-angle view of a data center aisle with glowing memory modules inside open server racks.

What CXL is in plain English


CXL is a connection standard for servers. It lets a processor communicate with memory and attached devices over a fast physical connection that is already common in computers.


That connection is based on PCI Express, often called PCIe. PCIe is the slot and signaling system used for hardware such as graphics cards, storage drives, and network cards. CXL builds on that base, then adds rules that help processors and memory devices work together more closely.


The important part is this: CXL allows a processor to access memory that is not sitting directly beside it on the main server board.


That external memory can come from a memory expansion card, a memory box, or in later designs, a shared pool of memory that several servers can draw from. The processor still sees that memory in a controlled, standard way.


The CXL Consortium, the industry group that manages the standard, has published several versions of the specification. Early versions focused on connecting a processor to attached devices. Later versions added features for switching and pooling, which are key for shared memory designs in large systems.


A simple way to think about it:


Traditional server memory

CXL-connected memory

Installed inside one server

Can be attached outside the main memory slots

Hard to change without replacing hardware

Can be added through CXL devices

Often stranded when a server does not need it

Can support more flexible allocation

Bound tightly to one processor

Designed for more flexible processor-to-memory setups


CXL does not make memory infinite. It also does not replace the fastest memory located right next to the processor. Local memory still matters because it gives the lowest delay. But CXL gives system builders another tier of memory, one that can be larger and more adaptable.


Why memory became the pressure point


For years, data centers solved many performance problems by adding more servers. That worked when software was easier to split across machines. Today, major workloads often hit memory limits before they run out of raw processing power.


Artificial intelligence is one reason. Large language models and recommendation systems can need huge amounts of memory to hold model data and active work. Databases are another example. The more data they keep in memory, the faster they can respond, because memory access is much faster than reading from storage.


The same issue appears in virtualization, where many smaller computing environments run on the same physical server. If memory runs short, the server may slow down even when the processor still has capacity left.


This creates a practical problem. Memory is expensive, power hungry, and often underused.


One server might be packed with memory for a peak workload that only happens at certain times. Another server might need more memory but have no easy way to get it. In many setups, the only choices are to buy a larger server, add more machines, or accept lower performance.


CXL changes that design question. Instead of asking, “How much memory should every server carry for its worst day?” engineers can ask, “How can memory be placed where it is needed, when it is needed?”


That idea sits behind composable memory, where memory is treated more like a resource that can be assigned, resized, and managed across systems rather than locked inside one box forever.


Samsung’s CXL strategy is about more than larger DRAM


Samsung is one of the world’s largest suppliers of dynamic random access memory, better known as DRAM. DRAM is the fast working memory used by servers, laptops, phones, and many other devices. CXL gives the company a way to sell memory in new forms, not only as chips that go into standard memory slots.


One visible part of the strategy is the CXL memory expander. This type of device adds more memory capacity to a server through a CXL connection. For workloads that need more memory but do not always need the very lowest delay, expansion can be useful.


Another part is software. Memory hardware is only useful if operating systems and data center tools can manage it well. Samsung has worked on software layers for CXL memory management, including tools meant to help systems place data in the right kind of memory. That matters because CXL memory may behave differently from local memory. Some data should stay close to the processor. Other data can live in expanded memory without a major slowdown.


A third part is ecosystem work. CXL only succeeds if processors, memory makers, server vendors, and software providers agree on how products should interoperate. Public support for CXL has come from major processor and server companies, and the CXL Consortium includes many large technology firms. That broad support is one reason CXL has drawn attention. A memory standard is much more useful when many companies build to it.


Samsung’s position is clear: if future servers need much larger and more flexible memory systems, memory suppliers must provide both the parts and the management layer around those parts.


Close-up view of a CXL memory expansion card with rows of black memory chips and a metal connector.

How CXL could change server design


CXL is not just about adding another card to a server. Its bigger promise is changing how memory is planned and used.


Servers can add memory without a full redesign


Traditional server memory depends on memory channels connected directly to the processor. That design is fast, but it limits how much memory a server can hold. If a workload needs more than the server supports, the options get expensive.


CXL memory expansion gives server makers another path. They can support larger memory footprints through attached devices. This is especially useful for workloads that need capacity more than the absolute lowest delay.


For example, an in-memory database may keep a “hot” set of frequently used information in local memory and less active data in CXL-attached memory. The application still gains capacity, while the system keeps the most time-sensitive data close to the processor.


Memory can be pooled across systems


Later versions of the CXL standard support switching and pooling. In a pooled setup, memory can sit in a shared device or shelf and be assigned to different servers as needed.


This could reduce stranded memory. Stranded memory is capacity that exists but cannot be used by another machine. It is common in large data centers because workloads rise and fall at different times.


If memory pooling matures, operators could size systems with less waste. A server running a memory-heavy job could receive more capacity. When the job ends, that capacity could return to the pool.


This will not happen perfectly or instantly. Shared memory must be secure, predictable, and easy to manage. Data centers also need tools that can monitor performance and prevent one workload from hurting another. Still, the direction is important.


Different kinds of memory can work together


Server memory design is moving toward layers. The fastest memory sits closest to the processor. Larger memory sits farther away. Storage sits even farther away but offers much more capacity.


CXL helps connect these layers in a common way. A system might use local DRAM for the most active data, CXL DRAM for added capacity, and storage for colder data. Future systems may mix different memory technologies as well.


The goal is not to make every layer identical. The goal is to place data where it makes the most sense for cost, speed, and power.


Why this matters for AI and cloud computing


AI workloads make the memory problem easier to see. Training and running large models requires moving large amounts of data between processors and memory. If the processor waits for memory too often, expensive chips sit idle.


Graphics processing units, often called GPUs, have their own high-speed memory. But many AI systems also depend on large pools of system memory to feed data, manage model pieces, or support many users at once. More flexible memory can help keep these systems better supplied.


Cloud computing has a similar need. Cloud providers run many workloads from many customers on shared hardware. Some use a lot of processor power. Some use a lot of memory. Some need both for only a few hours.


Fixed hardware makes this hard to balance. If a cloud provider buys servers with very large memory footprints for every case, many machines will waste capacity. If it buys smaller machines, some workloads will not fit well.


CXL offers a middle path. Memory can become a more adjustable part of the system, closer to how cloud platforms already treat processing and storage.


That does not mean every cloud server will suddenly use shared memory pools. Hardware refresh cycles take years. Software support takes time. Operators must prove that performance and security meet their standards. But the pressure from AI, analytics, and large databases gives CXL a strong reason to exist.


The challenges CXL still has to solve


CXL has strong industry support, but it is not magic. Several hard problems remain.


Performance must be predictable. Local memory is still faster because it sits close to the processor. CXL-attached memory adds distance and extra communication steps. For some workloads, that added delay is fine. For others, it can hurt.


Software must understand memory tiers. Applications and operating systems need better ways to decide what data belongs in local memory and what can move to CXL memory. If software treats all memory as the same, performance may suffer.


Security must be built in. Shared memory creates questions about isolation. One server should not read another server’s data. Hardware and software controls must make those boundaries clear.


Standards must stay compatible. CXL’s value depends on products from different vendors working together. The standard has to evolve without breaking trust in real deployments.


Costs must make sense. Extra hardware, switching equipment, and management tools all add cost. CXL must save enough money, improve enough performance, or allow enough capacity to justify that cost.


These challenges are normal for a new server architecture. The same kind of gradual adoption happened with solid-state drives, high-speed networking, and hardware accelerators. Data centers tend to adopt new hardware when the benefit is clear and the risk is manageable.


What to watch next


The future of CXL will be decided less by press releases and more by deployment patterns. The useful signals will be practical ones:


  • More server platforms with CXL memory support

  • Operating system features that manage tiered memory well

  • Memory expanders tested with real database, AI, and analytics workloads

  • CXL switches and pools used beyond lab trials

  • Clear guidance on when CXL memory helps and when local memory is better


The most likely near-term role for CXL is memory expansion for specific workloads. Pooling across many systems may take longer because it requires more changes in hardware, software, and operations.


For Samsung, the opportunity is clear. If servers start using CXL memory at scale, memory suppliers can move from selling fixed capacity to providing systems that help data centers place memory more carefully. That includes DRAM, controllers, modules, and software.


The bigger point is that CXL turns memory into a design choice instead of a fixed limit. It helps data centers answer a simple but costly question: how can large amounts of memory be used where they matter most?


Composable memory matters because modern computing no longer has one neat shape. AI, cloud services, databases, and analytics all pull memory in different directions. CXL gives the industry a shared way to respond, and Samsung’s strategy shows how seriously memory makers are preparing for that shift.


bottom of page