SK hynix has proposed a new AI data center design, the Memory-Centric AI Rack System (MARS). The company unveiled it at the OCP Global Summit 2026 to address memory bottlenecks that limit performance in processor-centric server racks.
MARS uses memory disaggregation, separating memory from CPUs and GPUs and combining it in large, shared pools. This allows memory resources to be allocated more flexibly, helping prevent some servers from having idle memory while others face shortages.
SK hynix says the approach can improve AI inference performance by up to five times. It can also reduce total cost of ownership by avoiding the need to overprovision GPUs.