GPU Platforms
SK hynix’s MARS Demos Bring Shared AI State Into the Rack Decision
SK hynix’s October 2026 MARS demos raise a rack-buying question: does an offer allocate memory, or support two hosts consuming the same AI state?
SK hynix’s October 16, 2026 report on OCP Global Summit, held October 12–15 according to the company, describes partner demonstrations of MARS, its Memory Centric AI Rack System. [1]
For a buyer, the consequential question is which relationship between hosts the offer supplies. Allocating more memory to two servers and enabling both to consume one common state object are different results. A memory-capacity line cannot settle whether the application gets the second result.
Read the demonstrations as separate evidence
The MARS section describes pooled-memory AI-agent serving with Arcee, Liqid and Marvell, and a separate pooled-memory streaming demo with Pickle. A distinct demonstration with AMD, Penguin, Xcena and Arcee uses CXL memory and DRAM/SSD hybrid memory, anticipating needed KV state and loading it into DRAM. Separately, the NAVER work is described as PoC-level validation of KV-cache and embedding sharing. [1]
HBF appears as a structural model and internally simulated performance graphs. That is a different evidence category from the system demonstrations. [1] These reports leave shipment availability and the buyer-specific supported delivery unresolved: the hardware, software and responsibilities that would make the intended state reuse work in the proposed environment.
Capacity allocation and common-region access answer different questions
The historical CXL 3.0 white paper supplies useful architectural vocabulary. On page 3, pooling means CXL-attached memory can be allocated and deallocated among hosts. Sharing adds hardware-coherent simultaneous access by multiple hosts to a common memory region. [2]
Pooling answers which host receives capacity. Common-region sharing answers which hosts can access the same backing region concurrently. Hardware coherency concerns visibility of the data held there; the application still needs to know what that data represents and when it is valid to consume.
This is historical architectural context, not a claim about the latest CXL edition. A standards feature establishes no support for a particular MARS demonstration or installed server, memory device, fabric, operating system and framework combination. The offer has to identify the relationship its actual combination supports.
The existing CXL host-memory expansion guide covers capacity, host-memory presentation and GPU locality. Here, the additional decision concerns the relationship between hosts consuming the state.
MARS-OCP-01: two consumers, two different payload counts
Consider two fictional alternative offers, each stipulating a 64 GiB memory pool and hosts A and B. Assume both invented implementations support their stated arrangements. Both consumers need the same read-only 16 GiB state object R, stipulated to be semantically eligible for both workloads. Assume R is completely published, has an agreed identity and valid generation, and remains available and unchanged throughout both uses. Both consumers have permission and a compatible state representation. These application conditions are supplied independently of the memory architecture.
Offer P assigns a separate 32 GiB region to each host and places one 16 GiB copy of R in each region. Two host consumers use two backing regions for R. The stipulated pool-backed R payload is 16 GiB + 16 GiB = 32 GiB.
Offer S explicitly supplies a supported implementation with one hardware-coherent 16 GiB common region holding R, simultaneously accessible to both hosts. Two host consumers use one backing region for R. The stipulated pool-backed R payload is 16 GiB. Other allocations in S are unspecified.
These are invented architectures, not two installed SK hynix products. The count includes only R’s payload in the stated pool regions. It excludes host caches, metadata, GPU copies or staging, code, allocation overhead and other reservations. No duration, bandwidth, power or price is supplied. The 32 GiB versus 16 GiB comparison establishes no observed total system memory saving, free capacity, latency advantage, additional GPU HBM or universal zero-copy behavior.
Both offers state the same pool capacity and support two consumers under the example’s assumptions. Their backing relationship for R differs. Crediting P with S’s single common region would add a capability the offer never supplied. Conversely, S’s stipulated common region does not establish that the complete application runs faster or costs less.
Connect the memory relationship to the application consumer
A useful offer should finish this sentence precisely: hosts A and B will consume the identified state through this supported arrangement. The supplier should distinguish separate allocations, an explicit state-transfer path and simultaneous access to one common region. Cross-host reuse can use other documented transfer or staging arrangements; their copies and access paths belong in the comparison. The phrase pooled memory does not choose among them.
For a common-region claim, request the actual host, memory-device and fabric identifiers, applicable firmware and operating-system versions, and the supported runtime or framework path. The evidence must cover the proposed relationship across those components. Listing each component as CXL-capable independently leaves their combined behavior unanswered.
The application-side consumer then needs an agreement about the state’s source, representation, valid generation, permissions and lifetime. It must identify which consumers may use that state and how they recognize the applicable object. A coherent view of bytes cannot establish that those bytes are eligible for this request or produce a correct final answer.
The prompt-cache versus answer-cache guide explains reuse eligibility and the work a cache hit represents. MARS-OCP-01 assumes that decision is already settled so the memory relationship can be compared on its own.
Name the integration provider responsible for joining the supported hardware path to that consumer, alongside the platform and application owners who supply the prerequisites. Ask for applicable evidence showing both intended hosts consuming the identified state under the agreed lifetime and access conditions. If one common region is promised, the evidence should identify that region rather than merely show two successful reads from unspecified allocations. Any application response or completion target needs its own evidence for the intended workload.
Resolve the supplied result before scoping a rack project
The October demonstrations broaden the architecture discussion by putting memory relationships alongside capacity in the rack decision. The unresolved customer-specific result is whether the proposed components and software supply the intended cross-host state use, with an accountable integration owner and applicable evidence. A buyer can use that requirement to establish what an existing fleet already supplies and what a proposed rack would actually add.
An evaluation using existing installed equipment may need no physical project. If the selected configuration separately requires mounting or physical connections, Leviathan’s physical installation service and its project inquiry provide a route to discuss the supplied equipment, approved configuration, permitted physical tasks and handoff records. [3]
01Footnotes
Footnotes
- SK hynix, OCP Global Summit report, October 16, 2026. Vendor account of separate MARS demonstrations, NAVER PoC work and the HBF simulation exhibit. Read October 20, 2026.
- CXL Consortium, Compute Express Link 3.0 white paper, page 3, Memory Pooling and Sharing and Figures 3–4. Historical architectural definitions; implementation support requires separate evidence. Read October 20, 2026.
- Leviathan Systems, GPU Infrastructure Physical Installation. Customer-supplied equipment mounting, approved interconnections, labeling and physical handoff records; tasks are engagement-defined. Read October 20, 2026.
Contact
Discuss Your Infrastructure Project
Tell us about your project, and our team will follow up to discuss the scope, requirements and availability.