Artificial intelligence is becoming more powerful every year, but there is a major problem slowing its progress: memory limitations.
Researchers at Electronics and Telecommunications Research Institute, also known as ETRI, say they have developed a new technology that could help solve this issue and make future AI systems faster, larger, and more efficient.
The technology is called “OmniXtend,” and it is designed to overcome what computer engineers often call the “memory wall.”
This problem happens when processors like GPUs become extremely fast, but the memory systems feeding them cannot keep up. As a result, expensive AI hardware spends too much time waiting for data instead of processing it.
This challenge has become especially serious as modern AI models continue growing in size. Large language models and advanced AI systems require enormous amounts of memory during training and operation. In many cases, adding more GPUs alone is no longer enough.
Traditionally, memory inside servers is tightly connected to individual machines. This means each computer can only use the memory physically installed inside it. Expanding capacity often requires replacing or upgrading entire systems, which can be extremely expensive.
ETRI’s new approach changes this idea completely.
Instead of keeping memory isolated inside individual servers, OmniXtend uses standard Ethernet networks to connect memory resources together across multiple machines. This creates what researchers describe as a giant shared “memory pool.”
In simple terms, servers and AI accelerators can access memory from other systems over the network as if it were part of their own hardware. This allows memory capacity to grow much more flexibly without constantly replacing equipment.
The researchers say this system also reduces delays caused by moving data between devices, helping AI training run faster.
Unlike older technologies such as PCIe, which have limits on connection distance and scalability, the OmniXtend system works with regular Ethernet switches commonly used in data centers. This makes it easier to build extremely large AI systems spread across many machines.
To make the technology work, the ETRI team developed several key components, including a memory expansion node based on FPGA technology and a special Ethernet-based memory transfer engine.
During demonstrations, researchers showed that multiple devices could successfully share and access memory across an Ethernet network in real time.
The team also tested the technology using large language model workloads similar to those used in modern AI systems. In situations where memory was limited, AI performance dropped sharply. But when memory was expanded through the Ethernet-based system, performance improved by more than two times.
According to the researchers, this suggests the technology could help maintain high AI performance without requiring huge amounts of expensive local memory inside every server.
ETRI recently presented OmniXtend at major international events including the RISC-V Summit Europe 2025 and RISC-V Summit North America 2025.
The institute says the technology could eventually be used in AI training systems, data centers, network switches, autonomous vehicles, ships, and future high-performance computing systems.
Researchers hope the breakthrough will help support the next generation of global AI infrastructure as AI models continue growing larger and more demanding.
