SK hynix在10月16日的OCP展会回顾中披露二代CMM-DDR5：256GB容量、CXL 3.2接口，并支持内存池化。厂商给出的带宽为80GB/s，称较上一代超过翻倍、能效提高约30%。原文未列完整测试配置与能效分母，这些只能作为厂商比较，不能换算成AI服务提速或整机省电幅度。

In its October 16 recap of the OCP event, SK hynix disclosed the second-generation CMM-DDR5, with 256 GB of capacity, a CXL 3.2 interface and support for memory pooling. The vendor reports bandwidth of 80 GB/s, more than twice the previous generation, and an improvement in energy efficiency of roughly 30%. The original article does not provide the complete test configuration or the denominator used for the efficiency comparison. These figures can therefore be treated only as vendor comparisons, rather than translated into AI-service acceleration or full-system power savings.

更值得关注的是系统演示：公司与NAVER Cloud验证了智能体场景的KV cache共享，以及多模态服务的嵌入数据共享。这里明确处于概念验证阶段，公告没有给出这套方案的普遍供货或量产日期。

The system demonstrations are more noteworthy: the company worked with NAVER Cloud to validate sharing of KV caches in agentic workloads and sharing of embedding data in multimodal services. These are explicitly proofs of concept. The announcement does not provide a date for general availability or volume production of the complete solution.

KV cache保存语言模型已经算出的注意力键和值。若多个服务确实需要相同且兼容的缓存，共享池有机会减少重复计算、复制和各服务器预留的闲置容量。可以把它理解为团队共用资料柜：价值取决于大家是否反复取用同一批资料，而不只是柜子有多大。

A KV cache stores the attention keys and values that a language model has already computed. If multiple services genuinely need the same compatible cache, a shared pool may reduce repeated computation, copying and idle capacity reserved on individual servers. It can be thought of as a shared filing cabinet for a team: its value depends on whether people repeatedly use the same materials, rather than merely on the size of the cabinet.

这带来一个架构选择：将可复用数据放在服务器之间的共享层，把紧迫的计算留给处理器。它并不直接证明GPU旁的HBM可以被替换；容量利用率、访问延迟与带宽是不同问题。

This creates an architectural choice: place reusable data in a shared layer between servers while leaving time-critical computation close to the processor. It does not directly establish that HBM beside a GPU can be replaced. Capacity utilization, access latency and bandwidth are different questions.

2023年Google研究者对CXL内存池提出过布线与交换成本、访问延迟、软件复杂度和实际可节约容量的质疑。这是架构背景反证，并非对本次新品的测试。对这轮演示同样应追问：缓存复用率多低时会失去优势？多机争用后，少数慢请求会不会拖累体验？仅看到模块带宽，回答不了这些问题。

In 2023, Google researchers raised concerns about CXL memory pooling, including cabling and switching costs, access latency, software complexity and the amount of capacity that could actually be saved. That work provides an architectural counterargument, rather than a test of this new product. The same questions should be asked of this demonstration: below what cache-reuse rate does the advantage disappear, and could a small number of slow requests impair the experience when multiple machines contend for access? A module's bandwidth alone cannot answer those questions.

下一步需要明确的供货与客户部署信息，以及公开模型、上下文长度、并发量、主机和网络基线的端到端测试。比较应同时报告首字响应、尾延迟、吞吐量与整机耗电，并计入共享设备成本。若低复用或高争用负载下收益消失，就应把适用范围收窄到特定服务，而非推广成所有AI机房的升级答案。

The next evidence should establish availability and customer deployments, along with end-to-end tests that disclose the model, context length, concurrency, host configuration and network baseline. Comparisons should report time to first token, tail latency, throughput and complete-system power consumption together, including the cost of shared equipment. If benefits disappear under low-reuse or high-contention workloads, the suitable use cases should be narrowed to particular services, rather than presenting the architecture as an upgrade for every AI data center.