SK hynix said on 16 October 2026 that it took part in OCP Global Summit 2026 in San Jose (12–15 October), presenting a “memory-centric” server-system vision built around CXL pooled memory, partner rack demos and compute-near-memory prototypes.
OCP Global Summit is the Open Compute Project’s annual hardware-standardisation conference for data-centre kit. SK hynix’s newsroom piece focuses on booth demos rather than a single product launch date.
Second-generation CMM-DDR5
At the show, SK hynix unveiled its second-generation CMM-DDR5 (256GB) module based on CXL 3.2. In the company’s wording, the product “more than doubles bandwidth (80GB/s)” versus the previous generation and “improves energy efficiency by about 30%.” Those bandwidth and efficiency figures are SK hynix’s own comparisons, not third-party benchmarks.
CMM (CXL Memory Module) is positioned as pooled memory that multiple servers can share over the CXL interface, so hosts draw capacity as needed. SK hynix says CXL pooled functionality is applied on this generation and that it has been working with NAVER Cloud since a September 2025 agreement on CXL-based bottleneck relief. At OCP it presented PoC-level validation of KV-cache sharing for agentic AI and embedding sharing in an omni-modal service called OmniServe — proof-of-concept language, not a general-availability claim.
MARS: Memory Centric AI Rack System
MARS (Memory Centric AI Rack System) is SK hynix’s name for a rack that integrates partner solutions with memory at the centre. The newsroom describes three demo tracks with partners:
- With Arcee, Liqid and Marvell: connecting servers and GPUs with pooled memory so KV-cache transfer/reuse improves versus network-only methods (company framing of the demo).
- With AMD, Penguin, Xcena and Arcee: a memory rack using CXL memory and CXL hybrid memory (DRAM + SSD), including predictive KV-cache loading and full-cache storage on SSD.
- With Pickle (personal AI agent provider): pooled-memory streaming meant to cut response latency across multiple models.
Treat the performance outcomes of these demos as vendor and partner demonstration claims.
CMM-Ax and long-context serving
CMM-Ax combines computation with memory so some work runs without moving data off the module, SK hynix says. A demo titled “Efficient Long Context LLM Serving” is meant to show acceleration of long contexts that are hard on conventional GPU-only setups. The company says the underlying research was published at MICRO 2026.
Broader portfolio on the floor
The Memory Section of the booth, per the same release, showed eSSD form factors (including DLC-oriented parts), DDR5 RDIMM capacities led by a 256GB 3DS RDIMM, MRDIMM, and 192GB SOCAMM2. An HBM area highlighted HBM4 / HBM4E alongside NVIDIA Vera Rubin partnership colour, and HBF (High Bandwidth Flash) as a high-capacity NAND path for inference-heavy agentic workloads. Those are exhibit and roadmap talking points from SK hynix’s newsroom, not shipping commitments dated in this article.
Vice President Hosik Kim’s executive session framed MARS as a response to the “memory wall” and rising power in GPU-centric racks, describing rack-level pooling and “dynamic zero-copy data sharing” over CXL.
Why it matters
Agentic and long-context inference inflate KV-cache and embedding footprints faster than many racks can move data over the network. CXL pooled memory and hybrid DRAM/SSD tiers are the industry’s main bet for sharing that state across hosts. SK hynix is using OCP — where hyperscalers align on open hardware — to show system-level demos, not only module datasheets.
TSN previously covered SK Hynix/Solidigm US NAND and China tool-licence reporting; this post is about the OCP memory-system story, not that fab thread.
What’s still unknown
- Commercial availability dates and pricing for second-gen CMM-DDR5 and CMM-Ax
- Which hyperscalers will take MARS-style racks beyond PoC
- Independent bandwidth and efficiency measurements versus SK hynix’s 80GB/s and ~30% claims
What to watch
Watch customer PoC expansions with NAVER Cloud and any CXL 3.2 module listings from OEMs after OCP. Also watch MICRO 2026 paper details on the long-context serving architecture.
What this means
- OCP demos, not a single GA launch. CMM-DDR5, MARS and CMM-Ax were shown at system level.
- Company comparisons. 80GB/s bandwidth and ~30% efficiency gains are SK hynix versus its prior generation.
- Partners on the rack. Arcee, Liqid, Marvell, AMD, Penguin, Xcena and Pickle appear in named demos.
Sources
Related stories
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
