CXL Memory Expanders
Cross-reference CXL memory expansion modules from Samsung, SK Hynix, and Micron for AI servers, HPC, and data centers
CXL (Compute Express Link) memory modules represent the next evolution in server memory architecture. Unlike traditional DIMMs that connect directly to CPU memory channels, CXL memory expanders plug into PCIe Gen5 slots (E3.S form factor) and extend memory capacity and bandwidth beyond what DIMM slots allow. CXL enables memory pooling — multiple servers sharing a common memory pool — and memory disaggregation — scaling memory independently from compute. These modules are critical for AI training, large language model inference, in-memory databases, and HPC workloads where memory capacity is the bottleneck.
60 GB/s
Samsung CMM-B (rack appliance)
24TB
Samsung CMM-B pooled
PCIe Gen5
E3.S form factor
CXL Memory Products
Samsung CMM-D MD220 128GB
MXFAH1280MB0-CCP
Capacity
128GB
Bandwidth
36 GB/s
Interface
PCIe Gen5 x8
Form Factor
E3.S (EDSFF)
Samsung CMM-D MD220 256GB
MXFAH2560MB0-CCP
Capacity
256GB
Bandwidth
36 GB/s
Interface
PCIe Gen5 x8
Form Factor
E3.S (EDSFF)
Samsung CMM-D 512GB
MXFAH5120MB0-CCP
Capacity
512GB
Bandwidth
36 GB/s
Interface
PCIe Gen5 x8
Form Factor
E3.S (EDSFF)
Samsung CMM-B Memory Pooling Box
Capacity
Up to 24TB
Bandwidth
60 GB/s
Interface
PCIe Gen5
Form Factor
Rack appliance (2U)
SK Hynix CMM-DDR5 96GB
Capacity
96GB
Bandwidth
~36 GB/s
Interface
PCIe 5.0 x8
Form Factor
E3.S (EDSFF)
SK Hynix CMM-Ax (CXL Memory-Accelerator)
Capacity
TBD
Bandwidth
TBD
Interface
PCIe 5.0
Form Factor
E3.S / Custom
Micron CXL 2.0 Memory Expansion 96GB
Capacity
96GB
Bandwidth
~36 GB/s
Interface
PCIe Gen5 x8
Form Factor
E3.S (EDSFF)
Micron H3 Falcon System
Capacity
Scalable (disaggregated)
Bandwidth
Scalable
Interface
PCIe Gen5
Form Factor
Rack-level system
Cross-Reference Matrix
Competing modules grouped by capacity tier. Use this to find equivalent products across manufacturers.
| Manufacturer | Product | Part Number | Capacity | Status | Notes |
|---|---|---|---|---|---|
| Samsung | CMM-D MD220 128GB | MXFAH1280MB0-CCP | 128GB | Production | Nearest Samsung equivalent (128GB vs 96GB) |
| SK Hynix | CMM-DDR5 96GB | — | 96GB | Sampling | Montage controller, 1a nm |
| Micron | CXL 2.0 96GB | — | 96GB | Production | Astera Labs Leo controller |
| Manufacturer | Product | Part Number | Capacity | Status | Notes |
|---|---|---|---|---|---|
| Samsung | CMM-D MD220 256GB | MXFAH2560MB0-CCP | 256GB | Production | Only option at this capacity |
| SK Hynix | — | — | N/A | Announced | No equivalent announced |
| Micron | — | — | N/A | Announced | No equivalent announced |
| Manufacturer | Product | Part Number | Capacity | Status | Notes |
|---|---|---|---|---|---|
| Samsung | CMM-D 512GB | MXFAH5120MB0-CCP | 512GB | Production | 3DS stacked DRAM |
| SK Hynix | — | — | N/A | Announced | No equivalent |
| Micron | — | — | N/A | Announced | No equivalent |
| Manufacturer | Product | Part Number | Capacity | Status | Notes |
|---|---|---|---|---|---|
| Samsung | CMM-B Pooling Box | — | Up to 24TB | Production | Up to 8 hosts, Supermicro certified |
| SK Hynix | — | — | N/A | Announced | CMM-Ax (compute, not pooling) |
| Micron | H3 Falcon System | — | Scalable | Production | 20× graph DB perf, Linux FAMFS |
CXL Technology Glossary
Key terms for understanding CXL memory architecture and products.
Compute Express Link — open standard high-speed interconnect built on PCIe physical layer, enabling cache-coherent communication between CPUs and memory/accelerators.
CXL Memory Module-DRAM (Samsung naming) — a CXL Type 3 device that provides DRAM-based memory expansion via PCIe Gen5 E3.S slot.
CXL Memory Module-Box (Samsung naming) — a rack appliance that houses multiple CMM-D modules and enables memory pooling across up to 8 servers.
CXL Memory Module-Accelerator (SK Hynix naming) — adds near-memory compute capabilities to CXL memory, enabling processing-in-memory for AI workloads.
EDSFF Enterprise and Datacenter Standard Form Factor — the physical slot used by CXL memory modules, the same slot as enterprise NVMe SSDs (formerly known as Ruler SSD).
First widely deployed CXL specification. Supports single-host attachment. Enables memory expansion but not pooling or switching.
Adds memory pooling, CXL switching, and multi-host support. Enables multiple servers to share a common CXL memory pool with dynamic allocation.
Adds multi-level switching, fabric capabilities, and enhanced peer-to-peer communication. Expected to reach broad deployment in 2026–2027.
Multiple hosts sharing a common pool of CXL-attached memory, with capacity dynamically allocated per workload demand. Reduces stranded memory across servers.
Decoupling memory from compute, allowing memory capacity to scale independently from CPU/server count. Enables right-sizing of memory per workload.
Non-Uniform Memory Access — CXL memory appears as additional NUMA nodes to the operating system, with slightly higher latency than local DRAM (~596 ns vs ~80 ns).
CXL device classification for pure memory expansion (no compute, just memory). All CXL memory expanders (CMM-D, Micron 96GB) are Type 3 devices.
