新川サーバー | エンタープライズ サーバー ハードウェア サプライヤー
The Most Underrated Three Letters in Enterprise Infrastructure
で 2026, every server you buy still follows the same architecture that has defined enterprise computing for 30 年: each CPU socket owns a fixed number of memory channels, and each channel supports a fixed number of DIMMs. If your workload needs more memory than the CPU can address, you buy another server — even if the first server has idle CPU cores sitting at 15% utilization. This is the “stranded memory” problem: a data center full of servers where either memory is full and CPU is idle, or CPU is maxed and memory is half-empty — but never both in balance.
CXL (Compute Express Link) changes this by decoupling memory from compute. A CXL 2.0 memory expander connects to the CPU via PCIe 5.0 lanes but presents memory to the operating system as if it were directly attached to the memory controller — not as block storage, not as a RAM disk, but as actual, byte-addressable system memory. This means a server with 16 DIMM slots can add terabytes of CXL-attached memory without adding a single DIMM slot or a second CPU socket.
How CXL 2.0 Works — And Why It Matters for Procurement
CXL runs over PCIe 5.0 physical lanes but uses a different protocol than NVMe or standard PCIe. Three CXL device types exist, but Type 3 (memory expanders) is the one entering production in 2026:
| CXL Type | Purpose | Production Status (2026) | Example |
|---|---|---|---|
| タイプ 1 (CXL.io + CXL.cache) | Accelerator cache coherency (GPU, FPGA sharing CPU cache) | Limited availability. NVIDIA, Intel working on implementations. | GPU with CXL cache-coherent memory access |
| タイプ 2 (CXL.io + CXL.cache + CXL.mem) | Accelerator with attached memory pool | Not yet production. Requires OS and hypervisor support not yet GA. | FPGA with shared memory pool for multiple hosts |
| タイプ 3 (CXL.io + CXL.mem) | Memory expander — adds capacity to host without adding DIMM slots | Production. Samsung, Micron shipping CXL 2.0 modules. Intel Xeon 4th/5th Gen with CXL 1.1 サポート. | Samsung 512GB CXL Memory Module, Micron CZ120 |
The practical impact: a Dell R770 with 32 DDR5 DIMM スロット (8 最大TB) could add another 2 TB of CXL-attached memory via two PCIe 5.0 x16スロット, delivering 10 TB of total system memory without moving to a 4-socket platform. For VMware clusters where memory is the first resource to run out, this means fewer servers, fewer CPU licenses, and lower power consumption — all from adding CXL modules to existing PCIe slots.
Which Servers Support CXL Today
| Platform | CXL Support | Max CXL Memory Per Server | Notes |
|---|---|---|---|
| Dell R660, R760, R770 (Xeon 4th Gen) | CXL 1.1 (タイプ 3 のみ) | まで 512 GB per CXL device, PCIe 5.0 x16 slot required | Requires BIOS update. Check CXL enablement in iDRAC. |
| Dell R670, R770 (ゼオン 6) | CXL 2.0 (タイプ 3, improved latency) | まで 2 TB per server (4バツ 512 GB CXL modules) | ゼオン 6 introduces native CXL 2.0 support with lower latency. |
| Dell R860 (4-socket Xeon 4th Gen) | CXL 1.1 | まで 2 結核 (4バツ 512 GB across available PCIe 5.0 スロット) | Already has 16 TB via DIMMs — CXL adds headroom beyond DIMM ceiling. |
| xFusion 2288H V7, 2488H V7 | CXL 1.1 (BIOS-dependent) | Similar to Dell R760/R860 equivalents | Verify CXL enablement in iBMC. Not all V7 platforms ship with CXL enabled by default. |
| Dell R7715 (AMD EPYC 9005) | CXL 2.0 (AMD native support) | まで 2 結核 | AMD’s CXL implementation is more mature than Intel’s 1.1. |
CXL vs Optane PMem vs DDR5 — The Memory Hierarchy Expands
| テクノロジー | Latency | Capacity per Module | Cost per GB (relative to DDR5) | Byte-Addressable | Persistent | Best for |
|---|---|---|---|---|---|---|
| DDR5 RDIMM (64 GB) | ~90 ns | 64-256 GB | 100% (baseline) | はい | いいえ (volatile) | Active working set, hot data |
| インテル Optane PMem 300 (512 GB) | ~350 ns | 128-512 GB | ~60% | はい | はい (persistent) | Large in-memory DB, buffer pool extension |
| CXL Type 3 メモリー (Samsung 512 GB) | ~250-400 ns (CXL 2.0) | 256-512 GB per module | ~40-50% | はい | いいえ (volatile) | Memory expansion for VMs, containers, in-memory analytics |
| NVME SSD (7.68 結核) | ~75,000 ns | 1.92-30.72 結核 | ~3% | いいえ (block) | はい | Bulk storage, cold data |
CXL memory sits between DDR5 and Optane PMem on the cost curve. It is cheaper per gigabyte than DDR5 and Optane, with latency comparable to Optane on CXL 2.0. For VMware hosts, Kubernetes nodes, and in-memory databases that are memory-constrained but not latency-sensitive at the nanosecond level, CXL memory is the most cost-effective expansion path in 2026.
Three Procurement Implications for Your Next Server Refresh
- Buy servers with more PCIe 5.0 slots than you think you need. Every CXL memory module consumes one PCIe 5.0 x16 (or x8) スロット. A server spec’d today with “just enough” PCIe slots for current networking and storage will have zero room for CXL expansion in 2027. The Dell R770 with 6 PCIe slots and 2 OCP slots is well-positioned; the R660 with 3 PCIe slots is not.
- Do not max out DDR5 DIMM slots on Day 1. If your R770 has 32 DIMM slots and you populate all 32 today with 128 GB DIMM (4 結核), you have zero room for CXL memory expansion because the platform’s memory address space is fully allocated. Leave DIMM slots empty — or populate fewer, larger-capacity DIMMs — and plan for CXL to fill the remaining memory capacity gap.
- Factor CXL into your 5-year TCO. A server that costs $40,000 today with 1 TB of DDR5 can grow to 3 TB of total memory (1 TB DDR5 + 2 TB CXL) over its 5-year lifespan by adding CXL modules in Year 2 また 3 — without replacing the server. The alternative is buying a larger server on Day 1 で $65,000. CXL changes the procurement math from “buy for the peak” に “buy for the baseline, expand with CXL.”
Source CXL-Ready Servers Through Xincuan
We configure Dell PowerEdge and xFusion FusionServer servers with CXL-capable PCIe 5.0 slots and BIOS settings pre-enabled for CXL memory expansion. 工場直送価格, 3-年保証, global shipping, and free consultation on CXL memory planning for your workload.
Request a CXL-ready server configuration
Explore: Dell PowerEdge Servers | xFusion Servers | インテル Optane PMem 300