Intel has strengthened its portfolio of AI accelerators for data centers.

On the 14th, Intel unveiled 'Crescent Island,' a new data center GPU optimized for AI inference workloads, at the '2025 OCP Global Summit.'

This product, with its high-capacity memory and energy-efficient design, is expected to contribute to the construction of next-generation AI infrastructure to respond to the era of real-time, global inference.

“AI is evolving from static learning to agent-driven, real-time inference,” said Intel CTO Sachin Katti. “To handle these complex workloads, heterogeneous systems that match the right silicon to the right task are essential.”

Intel's Xe architecture-based GPUs deliver efficiency and scalability to meet the surge in token throughput.

Looking at the key features of the Crescent Island GPU, it is based on the Xe3P microarchitecture, optimized for performance per watt (PPW), and is capable of processing large amounts of data with 160GB LPDDR5X memory.

It also supports various data types suitable for a 'Token-as-a-Service' environment and an air-cooled enterprise platform.It has a power and cost-efficient design optimized for EZ servers.

Intel supports inference-centric workloads through integrated system design from hardware to orchestration.

The open, integrated software stack under development for the Arc Pro B series GPUs enables early optimization and iteration, providing a developer-friendly environment.

Intel is building end-to-end solutions spanning AI PCs, data centers, and the industrial edge, powered by Xeon 6 processors and GPUs, and is collaborating with the Open Compute Project (OCP) to enhance performance, energy efficiency, and developer continuity.

Crescent Island GPUs are scheduled to begin customer sampling in the second half of 2026, and more information about the product will be available at Booth B3 in Expo Hall at the 2025 OCP Global Summit.