SK hynix Unveils MARS and HBF: Up to 5x Faster AI Inference with 14x the Capacity of HBM
nashnova research
SK hynix launched two new memory technologies at the OCP Global Summit — MARS, a memory-centric rack architecture, and HBF, a high-bandwidth flash standard — targeting the growing memory bottleneck in AI inference and training. Inference time-to-first-token improves 3–5×; storage capacity reaches 14× that of HBM. The signal: memory, not compute, is becoming the next contested layer in AI infrastructure.
Why is AI inference hitting a memory wall?
Today's data-center architecture ties memory to individual servers. Idle memory on one machine cannot be shared with other nodes.
Adding memory usually means buying an entire new server node. This means → operators are forced to over-provision GPUs, driving up total cost of ownership (TCO).
In plain terms = memory is the bottleneck, but the fix has been buying more GPUs — spending on the wrong side of the problem.
How does MARS fix this?
MARS — Memory-Centric AI Rack-Scale System — disaggregates memory: it pulls memory out of individual CPU/GPU servers and pools it into a shared, rack-scale memory fabric.
In multi-turn AI-agent inference, the KV cache — temporary data the model uses to remember context — balloons rapidly. MARS stores that cache in the shared pool. Compared with recomputation, time-to-first-token (TTFT) improves 3–5×, with especially sharp gains in tail latency.
This means → MARS does not replace HBM (high-bandwidth memory). It raises overall memory utilization, reducing the need for GPU-side memory expansion and marginally suppressing HBM demand.
What is HBF, and how does it differ from HBM?
HBF — high-bandwidth flash — has a straightforward positioning: roughly half the bandwidth of HBM, but 14× the capacity.
In plain terms = HBF trades speed for space. When a large model must hold massive parameters and context for inference, "how much fits" matters more than "how fast it moves."
This reflects a shift toward tiered storage: HBM handles the fastest layer, HBF fills the capacity layer. The two complement rather than replace each other.
How far away is commercialization?
SK hynix released OCP HBF spec version 0.7 in July 2026, expects version 1.0 in early 2027, and plans to deliver HBF Cube samples in early 2028.
Full-scale commercial deployment is still roughly two years out, but the steady spec cadence signals that industry standardization is accelerating — this is past the slideware stage.
How quickly Samsung, Micron, and other rivals engage with the HBF ecosystem will be the key indicator of whether competition in this new memory tier begins ahead of schedule.
市场有风险，内容仅供研究参考，不构成投资建议。