Blog
SK Telecom and Panmnesia to Develop Next-Gen CXL-Based Data Center Architecture
As large-scale AI services continue to expand, data centers are investing heavily in massive deployments of high-performance GPUs, resulting in astronomical costs. Recognizing the need for sustainable scalability, SK Telecom and Panmnesia are focusing beyond simple GPU expansion to technologies that enable more efficient utilization of existing computing resources. Through this collaboration, the two companies aim to simultaneously improve cost efficiency and performance by innovating data center interconnect architecture based on Compute Express Link (CXL) technology.
Background: Limitations of Modern AI Data Center Architectures
Modern AI data centers typically configure servers with fixed ratios of CPUs, GPUs, and memory. Multiple servers are connected via networks to form racks, and multiple racks are interconnected to build data centers. However, as AI models become increasingly diverse and larger in scale, this architecture faces limitations in terms of cost-to-performance efficiency.
To address these challenges, the two companies propose:
- Breaking away from rigid, monolithic server architecture.
- Replacing traditional network-based interconnects with CXL
Challenge #1: Resource Inefficiency from Fixed Server Configurations
In conventional AI data centers, CPUs, GPUs, and memory are statically bundled within individual servers. As a result, unused resources in one server cannot easily be utilized by others. In particular, when memory capacity becomes insufficient, additional GPUs—often unnecessary—must be deployed alongside it, creating inefficiencies. This structure lowers GPU utilization rates and increases both capital and operational expenditures.
To solve this issue, SK Telecom and Panmnesia propose a disaggregated architecture in which computing resources are separated by type and flexibly composed as needed. Instead of being confined within servers, CPUs, GPUs, and memory are interconnected at the rack level through a CXL Fabric Switch, operating as a unified system. By dynamically allocating only the resources required for each AI workload, this approach minimizes unnecessary resource waste and maximizes cost efficiency.
Challenge #2: Performance Degradation from Network Overhead
The companies will also improve computational efficiency by fundamentally changing the interconnect mechanism. In conventional AI data centers, GPU collective operations—essential for large-scale AI training and inference—rely on general-purpose networks such as Ethernet. This process introduces data copies and software intervention, resulting in performance degradation.
To address this limitation, SK Telecom and Panmnesia will eliminate network involvement in computational paths and transition to CXL. By utilizing CXL, it is able to interconnect resources without traversing conventional networks.
At the core of this architecture is the Link Controller, an electronic component that can be integrated into CPUs, GPUs, AI accelerators, and memory devices. Within each device, it enables direct communication over CXL, replacing data transfer that previously required multiple data copies into simple memory access operations. Furthermore, the architecture enables GPU-to-GPU and GPU-to-memory communication without software intervention, significantly improving processing efficiency. As a result, AI data centers can deliver higher performance without increasing the number of GPUs.
Collaboration Details
Under this collaboration, SK Telecom will lead the design of an architecture optimized for real-world deployment, leveraging its large-scale AI data center construction and operational expertise, along with its experience in AI model development and commercialization.
Panmnesia will implement a CXL-Based AI Rack by applying its link solutions—including CXL Fabric Switches that serve as the core of physical connectivity and Link Controllers responsible for logical integration. Through this approach, the link architecture—previously confined within individual servers—will be extended beyond server boundaries to the rack level and above.
The two companies plan to validate the next-generation AI data center architecture by running real AI models and comprehensively evaluating GPU and memory utilization, latency, and throughput by the end of this year. Following this, they intend to conduct proof-of-concept deployments in large-scale AI data center environments and pursue commercialization and business expansion.
Availability
Panmnesia’s partners can request CXL Fabric Switches (including PCIe 6.4/CXL 3.2 switch samples) and Link Controllers (PCIe 6.4/CXL 3.2 controllers) utilized in this collaboration project. Link Controllers are available either as IP or as custom silicon solutions.
Panmnesia is advancing toward deployment readiness beyond the prototype stage by conducting long-duration operational testing in real-world AI computational environments to verify data transmission stability and interoperability.
Companies incorporating Panmnesia’s link technology into their CPU, GPU, AI accelerator, and memory devices are expected to further strengthen their competitiveness in the data center market by establishing system-level integrated reliability that extends beyond validation at the individual device level
