At OCP Global Summit 2026, SK hynix senior vice president Hoshik Kim described MARS as a rack-scale memory-pooling architecture and HBF as a separate NAND-based memory technology, according to TrendForce’s account of his remarks. The distinction is simple: MARS organizes how compute systems share memory; HBF is a type of memory.
MARS pools memory across a rack
MARS stands for Memory-Centric AI Rack-Scale System. It is a disaggregated architecture: compute and memory are separated, with memory pooled as shared rack-scale infrastructure instead of being confined to individual servers. That makes MARS a system design, not a memory medium.
The workload behind MARS’s reported TTFT result
Kim described a multi-turn AI-agent inference service that stored and used KV cache in shared memory. The KV cache holds key-value data reused during model inference. For that service, Kim reported a 3–5× improvement in time-to-first-token (TTFT) compared with recomputation, and said tail latency was reduced. The reported figure applies to that workload and that comparison.
HBF’s reported capacity and bandwidth
High Bandwidth Flash (HBF) is a NAND-based memory technology. Kim characterized its capacity as about 14 times HBM’s and its bandwidth as typically about half of HBM’s. HBM stands for high-bandwidth memory.
| Measure | HBF relative to HBM, as Kim described it |
| Capacity | About 14× |
| Bandwidth | Typically about half |
For flash memory generally, IndexBox’s technical overview describes slower writes and limited endurance as reasons HBF is aimed at read-heavy inference, where model weights remain relatively static, rather than write-intensive training. This gives HBF a distinct workload context alongside its capacity and bandwidth profile.
HBF within SK hynix’s AI NAND portfolio
MK reports that SK hynix groups its AI NAND offerings into AINP for performance, AIND for capacity and AINB for bandwidth, identifying AINB as HBF. The categories place HBF within a broader portfolio rather than making it synonymous with all of SK hynix’s AI NAND products.
The HBF development roadmap
Kim said the Open Compute Project’s HBF specification version 0.7 had been released in July and expected version 1.0 in early 2027. He also said SK hynix was developing an HBF concept prototype and planned HBF Cube sampling for early 2028.
