What Is CXL (Compute Express Link)? Memory Pooling Explained
CXL lets servers pool, expand and share memory across CPUs and accelerators over standard PCIe links — here is how the open standard actually works.

CXL, short for Compute Express Link, is an open interconnect standard that lets a server's CPU and the memory, GPUs, and accelerators attached to it share a single, cache-coherent pool of memory instead of each device being stuck with only the memory soldered or plugged directly onto it. It runs on top of the same physical cabling and connectors as PCI Express, but adds a coherency layer so a CPU and a device can both read and write the same memory address space without the complex software workarounds older designs needed. The result, once enough of the ecosystem catches up, is memory that can be pooled across multiple servers, expanded without buying a whole new motherboard, and shared between a CPU and its accelerators — which is exactly what AI training and inference clusters, running short on memory bandwidth and capacity, badly need.
Key facts: CXL is developed by the nonprofit CXL Consortium, whose board includes Intel, AMD, Nvidia, Samsung, SK Hynix, Google, Meta, Microsoft, Arm, Huawei, Cisco, HPE, and Alibaba. The standard reached version 4.0 in November 2025, doubling link speed to 128 GT/s. CXL runs over the PCIe physical layer but adds its own coherency protocols, CXL.io, CXL.cache, and CXL.mem. It is not a replacement for Nvidia's NVLink or the open UALink standard — it solves a different problem, CPU-attached memory, rather than accelerator-to-accelerator links.
What Is CXL (Compute Express Link)?
Compute Express Link is an industry standard for connecting a CPU to memory, accelerators, and smart I/O devices while keeping everything "cache-coherent" — meaning the CPU and the attached device both see a consistent, up-to-date view of shared memory, without software having to manually copy data back and forth or flush caches to stay in sync. Intel originally developed CXL and in March 2019 handed it to a new nonprofit, the CXL Consortium, which was formally incorporated that September. The consortium has since absorbed two rival interconnect efforts, Gen-Z and OpenCAPI, folding their member companies and technology into a single standard.
The core idea is simple to state and hard to build: PCI Express already gives devices a fast, standardized physical connection to a CPU, but it was designed for devices that manage their own memory and talk to the CPU through driver calls, not for two processors sharing one memory space. CXL keeps PCIe's cables, connectors, and electrical signaling (it runs on the PCIe 5.0 physical layer through CXL 2.0, and PCIe 6.0's PAM-4 signaling from CXL 3.0 onward) but adds a cache-coherent protocol stack on top, so a GPU or memory-expansion card can plug into a standard PCIe slot while behaving, from the CPU's point of view, almost like locally attached memory.
How CXL Works Over PCIe
A single CXL link actually multiplexes three distinct sub-protocols, each handling a different job:
- CXL.io is essentially PCIe itself — it handles device discovery, configuration, interrupts, and ordinary non-coherent reads and writes.
- CXL.cache lets an accelerator (a smart NIC, a custom AI chip) coherently cache and access the host CPU's memory through a low-latency request/response channel.
- CXL.mem is the flip side: it lets the host CPU coherently access memory that physically sits on the attached device, whether that is ordinary DRAM or persistent flash-backed memory.
In CXL 1.1 and 2.0, data moves across the link in 68-byte "flits," most of which carry payload with a small slice reserved for routing and error-checking. CXL 3.0 moved to a larger 256-byte flit format running over PCIe 6.0's PAM-4 signaling, which both doubled raw bandwidth and, in a "latency-optimized" mode, trimmed a few nanoseconds of latency by shrinking the error-checking overhead. CXL memory access still adds roughly 200 nanoseconds of latency versus memory plugged directly into a CPU's own memory controller — noticeable, but small enough that many workloads don't care, especially for colder data that would otherwise have to come from a network-attached memory server or NVMe.
CXL Device Types: Type 1, 2 and 3
The CXL spec defines three device categories, and which protocols a device uses depends on what kind of device it is:
- Type 1 devices (CXL.io + CXL.cache) are typically accelerators like smart NICs that have no memory of their own and rely entirely on the host's memory, accessed coherently.
- Type 2 devices (CXL.io + CXL.cache + CXL.mem) are general-purpose accelerators — GPUs, custom ASICs, FPGAs — that carry their own local memory but also need coherent access to host memory and vice versa. This is the category that matters most for AI accelerators.
- Type 3 devices (CXL.io + CXL.mem only) are memory-expansion and persistent-memory boards: cards whose entire purpose is to give the host CPU more addressable memory over a PCIe slot. Micron's CZ120 and Samsung's CXL memory modules, discussed below, are Type 3 devices.
Memory Pooling vs. Sharing vs. Expansion
These three terms get used loosely in marketing, but the CXL spec defines them as genuinely different capabilities, introduced at different points in the standard's evolution:
- Memory expansion is the simplest case and has existed since CXL 1.1: a Type 3 card adds more memory capacity to a single host, reachable over PCIe instead of the CPU's own DRAM channels, letting a server carry far more memory than its DIMM slots alone would allow.
- Memory pooling, introduced in CXL 2.0, lets a CXL switch divide a bank of memory devices into segments and dynamically assign those segments to different host servers, so a rack's worth of memory capacity can be allocated to whichever server needs it most at a given moment, rather than being permanently nailed to one machine.
- Memory sharing, introduced in CXL 3.0, goes a step further: multiple hosts or devices can coherently access the same memory segment at the same time, rather than each getting its own carved-out slice. This is what makes multi-GPU and multi-node AI workloads able to share large in-memory datasets without copying them.
CXL Generations: 1.1 Through 4.0
CXL has moved through several major revisions since its 2019 debut, each adding capabilities that build toward full rack-scale memory disaggregation:
| Version | Released | PCIe PHY | Peak ×16 bandwidth | Key additions |
|---|---|---|---|---|
| 1.0 / 1.1 | 2019 | PCIe 5.0 | ~63 GB/s | Basic cache-coherent host-to-device memory access |
| 2.0 | Nov 2020 | PCIe 5.0 | ~63 GB/s | CXL switching, memory pooling across hosts, device integrity & encryption |
| 3.0 | Aug 2022 | PCIe 6.0 (PAM-4) | ~121 GB/s | Memory sharing, multi-level switching, mesh/ring fabrics, Global Fabric Attached Memory (GFAM) |
| 3.1 / 3.2 | 2023–2024 | PCIe 6.0 | ~121 GB/s | Hot-page monitoring for tiered memory, firmware activation, management improvements |
| 4.0 | Nov 2025 | PCIe 7.0-class | ~242 GB/s | Doubled link rate (128 GT/s), bundled ports, added RAS features, zero added latency |
The jump from 2.0 to 3.0 is the one that matters most for AI infrastructure, since it's the point where memory sharing and non-tree (mesh and ring) fabric topologies arrived, letting a CXL fabric scale out to thousands of nodes via Port Based Routing rather than being limited to a simple host-and-devices tree.
Why CXL Matters Right Now
AI training and inference clusters are memory-hungry in a way conventional servers never were: large language models need huge amounts of fast memory to hold model weights and the key-value caches generated during inference, and GPU memory capacity has historically grown more slowly than the models that need it. CXL gives data-center operators a way to add memory capacity independently of CPU sockets, pool unused memory across a rack instead of leaving it stranded on idle servers, and tier "colder" data out to CXL-attached memory rather than paying for premium on-package HBM for everything. That's especially valuable against the backdrop of the current DRAM and HBM supply crunch, where memory makers are prioritizing AI customers and prices for conventional memory have risen sharply — CXL doesn't fix a shortage, but it does let operators get more usable memory out of the capacity that does exist by eliminating the stranded, underused memory that sits idle on individual servers.
CXL vs. NVLink vs. UALink: What's the Difference?
These three names show up together constantly, but they solve different layers of the same problem and are largely complementary rather than competing head-on:
| Standard | What it connects | Owner / governance | Primary job |
|---|---|---|---|
| CXL | CPU ↔ memory & accelerators | Open, CXL Consortium (Intel, AMD, Nvidia, Samsung, SK Hynix, others) | Memory expansion, pooling and sharing |
| NVLink | GPU ↔ GPU, GPU ↔ CPU | Proprietary, Nvidia | High-bandwidth scale-up links inside a GPU server/rack |
| UALink | Accelerator ↔ accelerator | Open, UALink Consortium (AMD, Intel, Broadcom, Google, Meta, Microsoft, others) | Open alternative to NVLink for scale-up accelerator pods |
In practice, a next-generation AI server rack can use all three at once: NVLink or UALink to link GPUs to each other at very high bandwidth inside a pod, and CXL to let those same GPUs and their host CPUs reach a shared, expandable pool of memory sitting outside any single GPU's own on-package HBM. UALink exists specifically because the industry wanted an open alternative to Nvidia's proprietary NVLink for GPU-to-GPU links, the same way CXL is an open alternative to proprietary memory-expansion schemes — but UALink and CXL address different connections in the same rack rather than competing for the same job.
Who's Building CXL
The CXL Consortium's board-level members include Intel, AMD, Nvidia, Samsung, SK Hynix, Arm, Google, Meta, Microsoft, Huawei, Cisco, HPE, Alibaba, and Astera Labs, with dozens of contributor-level members ranging from Micron, Qualcomm, and Rambus to Dell, Lenovo, Oracle, IBM, and Kioxia. That roster is notable mainly for who it includes despite being fierce rivals elsewhere: Intel, AMD, and Nvidia all sit on the same board for CXL even as they compete directly on GPUs, CPUs, and AI accelerator platforms like Nvidia's Vera Rubin, because none of them wants to fragment the memory-expansion market the way earlier interconnect efforts splintered.
Real CXL Products Available Today
CXL has moved well past the whiteboard stage. Samsung introduced a 512GB CXL memory module in 2022 using its own ASIC CXL controller, built on the EDSFF (E3.S) form factor and PCIe 5.0, claiming roughly four times the capacity and a fifth of the latency of its earlier CXL memory products. Micron's CZ120 memory-expansion modules, built with Intel and sampling since 2023, ship in 128GB and 256GB capacities over a PCIe 5.0 x8 link in the same E3.S form factor, aimed at exactly the kind of memory-expansion use case Type 3 devices were designed for. On the processor side, Intel's Xeon "Sapphire Rapids" and AMD's EPYC "Genoa" and "Bergamo" chips added CXL 1.1 support starting in 2021-2023, and both companies' newer server platforms have continued expanding CXL 2.0 support since. Operating systems have followed: Windows Server 2022 added native CXL memory-device support, and the Linux kernel gained CXL support with version 6.5.
Current Limitations and Adoption Hurdles
CXL's biggest practical hurdle isn't the technology itself but the chicken-and-egg problem of ecosystem readiness: pooling and sharing only pay off once CPUs, switches, and enough CXL-aware software all support the same CXL generation, and most production deployments today are still using simpler CXL 1.1/2.0 memory expansion rather than the more ambitious 3.0-class pooling and sharing. The roughly 200-nanosecond latency penalty versus local DRAM also means CXL memory works best as a capacity tier for less latency-sensitive data rather than a wholesale replacement for a CPU's directly attached memory. As CXL 3.x and 4.0 hardware reaches the market and operating systems and hypervisors get better at automatically tiering data between local and CXL-attached memory, analysts expect the technology to move from a niche expansion tool into a standard part of how AI data centers are built.
What's Next for CXL
CXL 4.0 specifications were finalized in November 2025, doubling the link's data rate to 128 GT/s with no added latency and adding reliability features aimed squarely at large-scale AI deployments; expect the first CXL 4.0-capable CPUs and memory-expansion cards to start sampling as the next generation of server platforms from Intel and AMD rolls out. Watch for memory makers including Samsung, SK Hynix, and Micron to keep pushing CXL memory-module capacities upward, and for cloud providers to start quietly offering CXL-pooled memory tiers as a cost-saving option alongside their standard instance types, much the way Nvidia's own Vera Rubin platform already leans on pooled and disaggregated memory for its largest AI training configurations.
Frequently asked questions
What does CXL stand for?
CXL stands for Compute Express Link, an open interconnect standard that lets CPUs, memory and accelerators share one cache-coherent pool of memory over a standard PCIe physical link.
Is CXL the same thing as PCIe?
No. CXL runs over the same physical cabling, connectors and electrical signaling as PCI Express, but adds a cache-coherent protocol layer (CXL.io, CXL.cache and CXL.mem) on top so a CPU and an attached device can share memory coherently, which plain PCIe cannot do.
What is the difference between CXL and NVLink?
CXL is an open standard for connecting CPUs to memory and accelerators and for pooling or expanding memory across a system. NVLink is Nvidia's proprietary high-bandwidth link for connecting GPUs to each other and to CPUs. They solve different problems and are typically used together in the same server rack rather than competing directly.
What is CXL memory pooling?
Memory pooling, introduced in CXL 2.0, lets a CXL switch divide memory devices into segments and dynamically assign them to different host servers, so unused memory capacity in a rack can be allocated to whichever server needs it instead of sitting idle on one machine.
Which companies support CXL?
The CXL Consortium's board includes Intel, AMD, Nvidia, Samsung, SK Hynix, Arm, Google, Meta, Microsoft, Huawei, Cisco, HPE and Alibaba, with dozens of contributor members including Micron, Qualcomm, Dell, Lenovo, Oracle and IBM.
What is the latest version of CXL?
CXL 4.0 was finalized in November 2025. It doubles the link's data rate to 128 GT/s compared with CXL 3.x, adds bundled ports and reliability (RAS) features, and remains fully backward compatible with CXL 3.x, 2.0 and 1.1.
Sources
- CXL Consortium — About CXLcomputeexpresslink.org
- CXL Consortium — CXL 4.0 Specification Releasecomputeexpresslink.org
- CXL Consortium — Member Companiescomputeexpresslink.org
- Samsung Newsroom — Samsung Introduces Industry's First 512GB CXL Memory Modulenews.samsung.com
- Wikipedia — Compute Express Linken.wikipedia.org
Mara Lindqvist edits the hardware desk. She covers graphics cards, processors, memory and storage, the foundries and chip designers behind them, and what the numbers on a spec sheet mean for people choosing a PC. Specifications in her stories come from manufacturer spec pages and datasheets.