CXL® Memory eXpander Controller

Overview CXL Technology Key Advantages Typical Application Scenarios Product List Resources Related News Contact Us

CXL® Memory eXpander Controller (MXC)

The explosive growth of data-intensive applications such as generative AI and large language models is placing unprecedented memory pressure on data centers—bandwidth shortages leave compute cycles idle, capacity constraints limit model scale, and traditional expansion methods are costly and inefficient. The key to unlocking AI's full potential lies in achieving efficient memory expansion, sharing, and pooling without system architecture redesign.

Montage Technology's CXL® Memory Expansion Controller (MXC) is purpose-built to address this challenge. Compliant with the CXL Type 3 specification, the MXC provides high-bandwidth, low-latency interconnect for CPUs and CXL-enabled devices, supporting memory pooling and sharing across hosts while simplifying the software stack and reducing total cost of ownership (TCO). Flexibly deployable in memory expansion cards (AICs), backplanes, or EDSFF memory modules, MXC dramatically scales memory capacity and bandwidth to meet the urgent demands of AI and other data-intensive applications.

Application of CXL MXC in Mainstream Server

CXL Technology

CXL (Compute Express Link) is a next-generation high-speed interconnect technology for data centers and AI servers. Built on the PCIe physical layer, it enables decoupling, expansion, and sharing of processor and memory resources through a cache-coherent interconnect protocol. For workloads such as generative AI and large language models, CXL breaks through the memory capacity and bandwidth limitations of traditional servers, supporting larger-scale model deployment, more efficient data access, and more flexible resource scheduling—thereby alleviating the "memory wall" problem and improving inference and training efficiency.

Break through memory capacity limitations to support large-scale model deployment

As LLM parameter counts scale ever higher, per-server memory capacity frequently becomes the main bottleneck in model deployment. CXL delivers additional memory capacity to servers via memory expansion devices, allowing the system to host larger model parameters, KV caches, and inference data — while minimizing performance penalties associated with model sharding and cross-node communication.

Achieve memory pooling and sharing to improve resource utilization

CXL supports the creation of shared memory pools, decoupling memory resources from individual servers and dynamically allocating them to compute nodes as needed. In generative AI clusters, this prevents memory shortages on some servers while others sit idle — boosting overall resource utilization and strengthening the system's ability to scale elastically.

Build a tiered memory hierarchy to optimize inference performance and cost

CXL collaborates with technologies like MRDIMM and HBM to create a tiered memory architecture. Hot data resides in high-bandwidth local memory, while large model parameters and cold data are offloaded to CXL expanded memory. This approach maintains high-performance access for critical data while dramatically increasing usable memory capacity — boosting generative AI inference throughput and lowering total cost of ownership.

Key Advantages

High Bandwidth, Low Latency

Compliant with the CXL Type 3 specification, providing high-bandwidth, low-latency interconnect for CPUs and CXL-enabled devices.

Memory Expansion & Pooling Support

Exceed per-server memory capacity, unify dispersed memory into a dynamic pool, support on‑demand scaling and cross‑node sharing, dramatically improving memory efficiency and system agility.

Flexible Deployment, High Scalability

Adapt to memory expansion cards (AICs), backplanes, or EDSFF memory modules, scaling memory capacity and bandwidth to meet the growing demands of AI and LLM workloads.

High Compatibility, Ecosystem Ready

Compatible with mainstream CPU architectures, standard DDR memory interfaces, and PCIe interfaces. Deployable without motherboard or firmware changes, lowering the barrier to migration and integration.

Performance Specifications

CXL Data Rate

Up to 64 GT/s
DDR Data Rate

Up to 8000 MT/s
Latency

Around 100 ns
Compliance

Included in CXL Consortium Integrators List

Typical Application Scenarios

Memory Expansion
When the scale of a large AI model or an in-memory database exceeds the memory capacity of a single server, traditional approaches often require replacing the server with higher-specification models or adding a greater number of servers—an approach that is both costly and inflexible. Montage Technology's MXC chip, deployed on memory expansion cards (AICs) or EDSFF memory modules, delivers terabyte-level memory capacity expansion without requiring changes to CPU or motherboard design. This enables users to meet the urgent demand for massive memory capacity in scenarios such as LLM inference and real-time data analytics on existing server architectures at a lower cost.

Memory Pooling
In cloud computing and multi-tenant environments, memory resources across servers are often difficult to share, resulting in wasted idle memory on some servers while others face severe shortages. Montage Technology's MXC chip supports CXL-based memory pooling across hosts, consolidating memory resources from multiple servers into a unified pool that can be dynamically allocated to different CPUs or accelerators on demand. This significantly improves overall memory utilization, reduces memory idleness, and simplifies the software stack, helping data centers substantially lower total cost of ownership (TCO).

Product List

Part No.	Description	CXL Spec	DDR Spec	Application	Package	Document
M88MX6852	Type 3 CXL Memory eXpander Controller	CXL 1.1/2.0/3.2; CXL 1 x8 port or 2 x4 ports; Speed up to 64 GT/s	JEDEC DDR5	Memory Add-In Cards; EDSFF Memory Modules	1211-ball FCBGA	Product Brief
M88MX5891/5851	Type 3 CXL Memory eXpander Controller	CXL 1.1/2.0; CXL 1 x8 port; Speed up to 32GT/s	JEDEC DDR4/DDR5	Memory Add-In Cards; EDSFF Memory Modules	767-ball FCCSP; 716-ball FCCSP	Product Brief

Resources

CXL Memory eXpander Controller Supporting Tools

CXL Memory eXpander Controller (MXC) Flyer

CXL Dynamic Capacity Device Technology Demo with Live Multi-Host System Using CXL 3.2 MXC Silicon