SK하이닉스는 HBM을 넘어서는 두 가지 기술을 공개했다. 최대 5배의 속도 향상을 제공한다고 밝힌 MARS와, 용량이 HBM의 14배에 달하는 HBF다.

SK Hynix is redefining the boundaries of memory architecture for AI infrastructure with two new technologies. At the 2026 OCP Global Summit, the South Korean storage giant unveiled its MARS memory-acceleration architecture and HBF high-bandwidth flash technology, which are aimed at the increasingly severe memory bottlenecks in today's AI inference and training workloads.

At the summit, Hoshik Kim, Senior Vice President at SK Hynix, said the MARS system can improve time to first token (TTFT) in multi-round AI inference scenarios by a factor of 3 to 5, while also significantly reducing tail latency. This performance gain addresses the resource imbalances inherent in conventional processor-centric cluster architectures and is expected to lower the total cost of ownership (TCO) associated with GPU over-provisioning.

HBF expands capacity to 14 times that of HBM while maintaining bandwidth close to HBM levels. Sample shipments are scheduled for early 2028.

The launch marks a departure by SK Hynix from the mainstream HBM track, forging a differentiated path. With HBM supply persistently tight and AI chipmakers such as NVIDIA driving surging demand for high-bandwidth memory, these technologies could serve as significant complements to existing AI storage architectures and have a profound impact on the competitive landscape of the memory industry.

**MARS: Breaking the GPU Over-Provisioning Dilemma Through Memory Disaggregation**

Traditional processor-centric distributed computing clusters face structural bottlenecks. Kim pointed out two fundamental shortcomings. First, resource imbalance: idle memory on one server cannot easily be shared with other nodes. Second, rigid scalability: adding memory capacity often requires acquiring an entire new server node. Together, these issues compel operators to over-provision GPUs, driving up TCO and reducing system efficiency.

The core logic of MARS (Memory-Centric AI Rack-Scale System) is memory disaggregation. Kim described it as "decoupling memory from the traditional CPU/GPU server paradigm and building a dedicated, large-scale memory pool," which enables a leap from "single-server memory surplus" to "pool-level efficiency." At the architectural level, MARS extends this tray-level memory-pool concept to the rack scale, allowing compute and memory to operate together as shared infrastructure.

MARS's value is especially pronounced in real-world applications. Kim cited a multi-turn AI agent reasoning service as an example, explaining that the system can store the rapidly expanding key-value cache from multi-turn dialogues in a shared-memory system. This avoids exhausting GPU memory and storage resources. According to on-site reporting by TrendForce, citing test data, Kim stated that compared with a recomputation-based approach, MARS can improve TTFT by 3 to 5 times, with especially notable reductions in tail latency.

For AI inference service providers, TTFT is a key metric for user experience, and tail latency reduction directly affects service stability under large-scale concurrent workloads. The GPU over-provisioning problem that MARS addresses is also one of the main drivers of the persistently high operating costs of modern AI data centers.

**HBF: Trading Bandwidth for Capacity, With Samples Expected in 2028**

HBF (High Bandwidth Flash) is another technology roadmap SK Hynix is pursuing alongside HBM. Kim clarified HBF's positioning relative to HBM: "HBF extends HBM-level bandwidth to rack-scale capacities… HBF's bandwidth is roughly half that of HBM, but its capacity is about 14 times greater."

This positioning highlights HBF's core value. By accepting a modest bandwidth discount in exchange for a substantial capacity increase, it enables a more resilient storage hierarchy at the rack level. In large-model inference, where massive numbers of parameters and context must be handled, capacity often constrains system performance far more than bandwidth does, so HBF's 14-fold capacity advantage has significant practical implications.

On the roadmap, SK Hynix released version 0.7 of the OCP HBF specification in July this year. Version 1.0 is expected early next year, and HBF Cube samples are slated for early 2028. That timeline suggests roughly two years remain before large-scale commercial deployment. The steady progress of the specification also points to an accelerating pace of industry standardization.

**Potential Impact on the AI Chip and Storage Industry Chain**

The launches of MARS and HBF reflect storage vendors' strategic push to diversify amid tight HBM supply. HBM production capacity is highly concentrated, and strong demand from AI chipmakers such as NVIDIA keeps supply under pressure, creating opportunities for alternative or complementary memory architectures.

From a technical standpoint, MARS does not directly replace HBM. By decoupling memory, it improves overall memory utilization at the system-architecture level and reduces reliance on expanding GPU memory capacity. Data center operators could therefore deploy MARS to improve inference performance without increasing HBM procurement, which may modestly dampen HBM demand. HBF, meanwhile, serves as a capacity supplement to HBM within the storage hierarchy, helping to ease capacity constraints in large-model deployments. Because its bandwidth is roughly half that of HBM, however, it is better suited as a complementary tier in a tiered storage architecture than as a direct replacement for HBM.

For the storage industry chain, SK Hynix's concentrated unveiling at this year's OCP Summit is a key signal to cloud computing vendors and AI infrastructure operators about its technology roadmap. As the HBF specification advances toward version 1.0 and moves into the sample stage in 2028, competition around next-generation storage architectures is likely to intensify sooner than expected. The response strategies of rivals such as Samsung and Micron warrant ongoing attention.