Company News · 采购简报

SK hynix CXL 池化内存与 MARS 机架演示，预示 AI 服务器 BOM 新选择

SK hynix 在 OCP Global Summit 2026 上展示了第二代 256GB CMM-DDR5、以内存为中心的 MARS 机架以及 CMM-Ax，并与 AMD、Marvell 和 NVIDIA 进行了实时 KV-cache 演示。

已发布 已更新 4 分钟阅读 ✓ AI 就绪 分享

TL;DR——三句话

  • SK hynix's second-gen CMM-DDR5 offers 256GB on CXL 3.2 with 80GB/s bandwidth and about 30% better energy efficiency than the prior generation (S1).
  • The MARS Memory Centric AI Rack System was demonstrated with AMD, Arcee, Liqid, Marvell, Penguin, Xcena, and Pickle, proving KV-cache reuse on pooled memory improves performance versus network-based methods (S1).
  • CMM-Ax combines computation with memory and was shown via an Efficient Long Context LLM Serving demo, with the research published at MICRO 2026 (S1).

摘自 key_takeaways （真实字段，最多 3 条）。事实与分析分别标注。

SK hynix CXL 池化内存与 MARS 机架演示，预示 AI 服务器 BOM 新选择 — Company News news briefEditorial cover for a Company News brief. Chip motif; headline "SK hynix CXL 池化内存与 MARS 机架演示，预示 AI 服务器 BOM 新选择". Reported key numbers: 256 GB; 80 GB/s. 1 cited source. A one-time entrance animation stops under reduced motion.COMPANY NEWS · NEWS BRIEFSK hynix CXL 池化内存与 MARS 机架演示，预示 AI 服务器 BOM新选择REPORTED KEY NUMBERS256 GB80 GB/s1 sourceXAIPART
Cover image generated from data already shown on this page: the category motif, the headline, and reported key numbers. No decorative illustration; no third-party image.

您从 AI 回答跳转而来——这是可验证的原文

AI 摘要可能遗漏了其结论的依据。原始预测、受影响料号及所有引用均在本页，均可追溯至来源。

XAIPART 预测

本篇简报包含可证伪的预测。

跳转至完整预测及其证伪规则 ›
受影响的料号

此报道在目录中无匹配——我们选择不显示，而不是推断匹配。

了解判定方式 ›
需要采购料号？发送请求 → 询价

Reported by SK hynix 官方新闻 on 2026-10-16. Summarized and analyzed by XAIPART Editorial.

What happened

SK hynix took part in OCP Global Summit 2026, held from the 12th to the 15th in San Jose, California, and demonstrated next-generation memory solutions at the system level, including a proposed memory-centric server architecture based on CXL pooled memory.S1

The System Demo section presented three solutions: CMM-DDR5, MARS, and CMM-Ax. CMM-DDR5 is a CXL Memory Module that lets multiple servers share a memory pool. SK hynix unveiled its second-generation product with 256GB capacity, based on CXL 3.2, more than doubling bandwidth to 80GB/s versus the prior generation and improving energy efficiency by about 30%.S1

MARS (Memory Centric AI Rack System) integrates partner solutions into a single rack with memory at its core, and three demos were shown. With Arcee, Liqid, and Marvell, an AI agent serving demo proved that transferring or reusing a KV cache on CXL pooled memory significantly improves system performance versus network-based methods. With AMD, Penguin, Xcena, and Arcee, a demo confirmed that a CXL hybrid memory rack can store large-scale KV caches on SSD while prefetching needed caches into DRAM. A demo with Pickle showed pooled-memory streaming reducing response latency in a personal AI agent service.S1

CMM-Ax, which combines computation with memory, was demonstrated through an Efficient Long Context LLM Serving system; the underlying research was published at MICRO 2026.S1 The Memory Section displayed eSSDs (PS1010 U.2, PS1110 E3.S, PS1101 E3.S, PEB210 E1.S, PEB210 M.2 2280 with DLC support) and server DRAM including a 256GB 3DS RDIMM, 128GB DDR5 MRDIMM, and 192GB SOCAMM2. Visitors could also mount HBM4 and SOCAMM2 onto an NVIDIA Vera Rubin model, and HBM4E, HBM4, and HBM3 products were exhibited alongside an HBF (High Bandwidth Flash) structure model.S1

Key numbers
MetricReported valueSource
CMM-DDR5 gen-2 capacity256GBS1
CMM-DDR5 gen-2 bandwidth80GB/sS1
Energy efficiency gain vs prior genabout 30%S1
CXL interface generationCXL 3.2S1
Pricing and availabilityNot disclosed

Background timeline

  1. - SK hynix developed the world's first HBM, according to the booth graphic described by the source.S1
  2. September 2025 - SK hynix signed a business agreement with NAVER Cloud to resolve bottlenecks in cloud AI services using CXL-based memory solutions; the source dates this to September of the year before the 2026 summit.S1
  3. - OCP Global Summit 2026 held in San Jose, California, where SK hynix presented MARS, CMM-DDR5, and CMM-Ax.S1

XAIPART take

Technical impact

The demos shift CXL memory pooling from datasheet claims to validated rack-level operation, which matters for KV-cache-heavy agentic AI serving. CMM-DDR5 doubling bandwidth to 80GB/s while adding a pooling function makes it a plausible substitute for some node-local DRAM expansion. CMM-Ax points toward in-memory compute for long-context inference, but no throughput figures were disclosed.

Market impact

Procurement teams should expect CXL 3.2 modules and memory-centric rack designs to enter hyperscaler qualification cycles within the next several quarters, since SK hynix has already validated PoC-level results with NAVER Cloud. The breadth of partner names (AMD, Marvell, NVIDIA, Liqid) suggests an emerging multi-vendor ecosystem, but no pricing or volume commitments were disclosed, so near-term purchasing should stay on standard DDR5 RDIMM and eSSD lines.

Action

Design-in teams planning 2027 AI servers should add CXL 3.2 pooled-memory slots to their platform RFPs and evaluate KV-cache tiering against hybrid DRAM-plus-SSD racks. XAIPART recommends requesting SK hynix sample availability and partner interoperability lists through distributor channels now, before volume allocation tightens.

XAIPART Forecast

XAIPART expects SK hynix's second-generation 256GB CMM-DDR5, based on CXL 3.2, to appear in at least one hyperscaler's published server reference design by Q3 2027; if no hyperscaler or OEM platform discloses CXL 3.2 pooled-memory support built on SK hynix CMM-DDR5 by 2027-09-30, this forecast is wrong. XAIPART analysis

Facts are as reported by the cited source. The design-in, sourcing and forward-looking statements above are XAIPART's analysis, not the source's claims. The report is single-source and vendor-authored; pricing, availability dates, and quantified demo performance were not disclosed.

Corrections: none at publication. This brief will be updated with a timestamp if the source corrects its report.

Key figures sharing one unit family (values are copyable from the table above).
Key numbers in the same unit familyComparable figures: CMM-DDR5 gen-2 capacity 256GB; CMM-DDR5 gen-2 bandwidth 80GB.CMM-DDR5 gen-2 capacity256GBCMM-DDR5 gen-2 bandwidth80GB

替代料与生命周期

已收录的替代料——仅为目录范围，不代表市场排名
料号关系生命周期依据
— 暂无收录替代料 待补充数据 — 需 match_type 后才可显示任何前代标签

“前代”和“替代”标签仅来自已记录的匹配类型、生命周期状态或来源文本。没有记录匹配时，不显示任何标签。我们自己的商品列表不构成市场参考。

关注要点

可证伪预测——阈值、时间窗口和当前状态。被推翻的预测仍保留可见。
论断阈值发布方状态
XAIPART expects SK hynix's second-generation 256GB CMM-DDR5, based on CXL 3.2, to appear in at least one hyperscaler's published server reference design by Q3 2027 no hyperscaler or OEM platform discloses CXL 3.2 pooled-memory support built on SK hynix CMM-DDR5 by 2027-09-30 Pending

被证明错误的预测仍保留可见并加以标注。我们不会删除它们。

受影响的料号

目录中无匹配。 此事件未映射到 XAIPART 当前上架的任何料号。我们不推断匹配，也不显示占位列表。

各维度成本压力

Source: FRED · Generated · reference output z and composite are algorithm outputs, not a fetch snapshot

Cost pressure by dimension — composite stateDiverging bars for three dimensions around a zero midpoint. Input Cost has no reading; Supply Demand z -1.4711, notable; Fx Energy z 0.4541, neutral. Composite index -7.6, band neutral.0投入成本no reading供需z −1.4711 · notable46 d汇率与能源z +0.4541 · neutral6 dcomposite −7.6 · neutral−100 · −40 · −15 · 0 · +15 · +40 · +100
Underlying values — copyable for verification.
DimensionFRED series IDUnitLagAs of
投入成本PCOPPUSDM, PALUMUSDM, PNICKUSDMUSD/mt96 d
供需PCU334413334413, IPG3344Sindex(1982=100) / index(2017=100)46 d
汇率与能源DEXCHUS, DCOILBRENTEU, DHHNGSPrate / USD/bbl / USD/MMBtu6 d

No free authoritative interface for: gold · silver · rare earth · polysilicon · lithium carbonate · market share — listed rather than approximated.

本文如何生成

过程证据——抓取、来源可靠性检查、模型与校验
本简报的审计记录
Source fetched SK hynix 官方新闻 · news.skhynix.com · 54,785 chars · fetch status success
Source sanity (S1) ok — passed before any model call (no s1_ block recorded)
Model / prompt glm-5.3-flash · prompt Blog Agent - News Summary (v2.8-final) · 22,429 tokens · 36.9s · run completed
Gate verdict hold · 5/6 applicable checks passed · spec v2.9 · C10 ratio 0.00605 · C10 rounds 0
Time contract 2 <time datetime> values in this brief — 0 violations (+08:00 or date-only)

上述每一行均由流水线生成，而非人工撰写。

编辑评分

确定性校验结果——纯 PHP 实现，判定过程不涉及模型
校验评分明细 6 项适用检查， 5项通过： 83 %. Hold items: C12 (historical truthfulness), C13 (falsifiable forecast) — sent to human review. Gate spec: v2.9 (S0 + S1 + C1–C20) · Source affinity (C20): 0.875. 适用检查已通过 5 / 6 = 83% · 判定： hold Hold items: C12 (historical truthfulness), C13 (falsifiable forecast) — sent to human review Gate spec: v2.9 (S0 + S1 + C1–C20) · Source affinity (C20): 0.875 是否发布由校验决定。独立审计评分仅供参考。

相关报道

发现错误？ 告诉我们哪里有问题 ——更正内容会与原文一同发布并记录在案。我们不会悄悄修改文章。

评论

成为第一个评论的人。

来源

  1. Memory-Centric Server Systems Are the Future: SK hynix Presents Its Next-Generation AI Infrastructure Vision at OCP Global Summit 2026 — https://news.skhynix.com/en/ocp-global-summit-2026/ · rss · 访问于

事实标注为 S1/S2，并可追溯至所链接的来源。标注为 分析 的内容是 XAIPART 自身的推理，与报道的事实分开呈现。来源链接会被访问核实，而非 nofollow ——署名是有意为之。