SK hynix is proposing shared memory for AI servers after partner tests showed remote DRAM cut the wait for a model’s first response, ChosunBiz reported, highlighting a capacity constraint in AI computing.
At the 2026 Intel AI Summit Seoul on Thursday, Nov. 12, the South Korean chipmaker presented a memory rack designed to add pooled capacity alongside existing server systems. The system would make an additional memory pool available to CPU and GPU servers without replacing the memory each server already uses, ChosunBiz reported. The report tied the proposal to growing AI data-processing demands and limited room to place memory beside GPUs.
The issue arises as a model builds a key-value (KV) cache, information it retains during a task or conversation. Longer exchanges can expand the cache until a GPU runs short of memory, forcing a system to clear stored data or calculate it again.
In repeated-inference trials with partners, remote CXL DRAM produced an average time to first token of 15.45 seconds; the external SSD result was 33.56 seconds, while cache eviction followed by recomputation took 80.76 seconds, ChosunBiz reported. Time to first token measures how long it takes a model to begin responding.
SK hynix Vice President Joo Young-pyo told ChosunBiz that pooling memory could improve utilization and scalability and lower system costs. The tests measure performance in partner trials; they do not establish results across deployed systems.
The proposal adds to SK hynix’s earlier OCP 2026 work on memory technologies for AI servers, which included approaches intended to help servers share data and handle AI workloads.
SK hynix is validating the technology with customers and partners while working with Intel at the system-design stage, ChosunBiz reported.
