As AI commercialization expands, the accuracy of AI services has a direct impact on corporate sales. Big tech companies have recently been increasing memory adoption explosively in the process of improving the accuracy of AI services by utilizing larger amounts of data and larger models, which has resulted in excessive spending, such as additional server and component purchase costs.

Domestic fabless startup Panesia announced on the 18th that it unveiled the world's first AI cluster with a CXL 3.1 switch at the OCP Global Summit, the world's largest data center infrastructure event held in California, USA on the 15th.

The OCP Global Summit will discuss how to build the ideal data center infrastructure, including how to address the cost-effectiveness issues of existing data centers. This year, more than 7,000 stakeholders, including many global companies, gathered to focus discussions on AI-oriented solutions.

On this day, Panesia attracted much attention by introducing an AI cluster equipped with CXL, a solution that dramatically improves the cost efficiency of AI data centers by utilizing the next-generation interface technology, CXL (Compute Express Link).
Panesia's CXL-equipped AI cluster is a framework built using Panesia's main products, the CXL 3.1 switch and CXL 3.1 IP. The CXL-memory node, which stores large amounts of data, and the CXL-GPU node, which processes machine learning operations at high speed, are connected through the CXL 3.1 switch.

If you want to expand memory, you only need to add memory and CXL devices for memory expansion. This saves you unnecessary expenses on purchasing other server parts, and saves on memory expansion costs.

During the show, Panesia presented a demo accelerating its latest LLM-based application, Retrieval-Augmented Generation (RAG), on its AI cluster. According to the official, when utilizing Panesia's CXL-equipped AI cluster, inference delay time can be shortened by about 6 times compared to existing storage/RDMA-based systems.

A Panesia official explained, “Global companies related to data centers, such as Google, Microsoft, Supermicro, and Gigabyte, who attended the OCP Global Summit event, showed a lot of interest.” He added, “In particular, many companies that provide servers expressed their desire to introduce Panesia’s CXL 3.1 switch chip, which is scheduled to be provided to customers in the second half of next year, into their server products.”

Practical data center-level memory expansion is possible by connecting hundreds of different devices into a single system through Panesia’s CXL switches that support the high scalability related features of the CXL 3.1 standard and connectivity to CXL devices such as CPUs, GPUs, and memory expansion devices. The CXL 3.1 switches are scheduled to be available to customers in the second half of next year.

He continued, “NVIDIA, AMD, and others have shown interest in the fact that they can activate CXL on GPU devices using Panesia’s CXL 3.1 IP,” and “We look forward to the future when Panesia’s CXL IP is embedded in the accelerator products of these companies.”

Meanwhile, Panesia presented design technology for implementing CXL on GPUs at the OCP Future Technology Symposium and was also invited as a speaker at the Memory Fabric Forum presentation, introducing the latest CXL technology for accelerating AI to experts around the world.