SK Telecom said on the 4th that at "MWC26," the world's largest mobile communications exhibition being held in Barcelona, Spain, it signed a memorandum of understanding (MOU) with Panmnesia, a computing resource consolidation corporations, to jointly develop a "CXL (Compute Express Link, a data consolidation standard)-based next-generation artificial intelligence (AI) data center (DC) architecture." It is moving to innovate AI DC architecture. As AI models become more advanced and memory demand surges, the plan is to improve performance and expense efficiency at the same time by changing how computing resources are consolidated instead of simply adding GPUs.
CXL is a data consolidation standard that organically consolidates data among CPUs, GPUs, and memory to enable ultra-high-speed, low-latency processing, allowing flexible expansion and use of computing resources that had been tied to server units. The core of this collaboration is to use CXL-based technology to raise AI processing efficiency without unnecessary equipment expansion, improving the economics of AI DCs.
Panmnesia is a domestic startup with global-level capabilities related to CXL. It provides various link semiconductors (communication semiconductors that streamline data movement) needed to build efficient AI DCs, including a fabric link switch (a device that manages data flow by consolidating multiple devices in the middle) and a link controller (a device that helps efficient data transmission between devices).
Existing AI DCs have a fixed architecture in which CPUs, GPUs, and memory are bound at the server unit. Because of this, even if a specific resource remains on one server, it has been difficult to use it on another server. In particular, when memory runs short, the inefficiency of having to add GPUs that are not actually needed has been repeated. This architecture lowers GPU utilization and increases AI DC construction and operation expense.
To solve these problems, the two companies will apply CXL-based technology to shift to an architecture that can flexibly consolidate and combine CPUs, GPUs, and memory, moving away from a fixed, server-unit architecture. They will expand the scope of resource consolidation, previously limited to inside a server, to the rack unit that bundles multiple servers, enabling selective use of needed resources.
Along with this, the two companies will also change how resources are consolidated. Until now, GPU collaborative computing in AI DCs has been conducted by exchanging data over general-purpose networks such as Ethernet, but in this process, data copying and software intervention occurred, creating a bottleneck that delayed speed. Collaborative computing is the process in which multiple GPUs share and combine computation results with one another, and it is essential for large-scale AI training and inference.
The two companies plan to run actual AI models and comprehensively verify GPU and memory utilization, latency, throughput, and more, then unveil the next-generation AI DC architecture by the end of this year. After that, they plan to proceed with commercialization and business development following demonstrations in a real large-scale AI DC environment.
Jeong Seok-geun, head of SKT AI CIC, said, "AI DC competitiveness depends on system optimization that includes memory and data flow beyond GPU performance competition," and added, "This collaboration will ease the structural bottleneck of the 'memory wall,' where data movement and supply cannot keep up even when compute performance rises, boosting both AI DC performance and economics."
Jeong Myeong-su, CEO of Panmnesia, said, "Next-generation AI infrastructure is driven not by the performance of individual equipment but by the 'architecture' created by various link semiconductors," and added, "Together with SK Telecom, we will present a standard model for a high-efficiency AI DC that will draw global market attention."