Intel announced a new data center GPU codenamed Crescent Island at the OCP Global Summit 2025. The chip is intended for AI inference applications. It combines higher memory capacity and energy efficiency in a design suitable for air-cooled servers.
Crescent Island is part of Intel’s renewed effort to gain a foothold in the AI accelerator market. In recent years, the company has lost market share to competitors such as Nvidia and AMD.
The GPU is based on the new Xe3P architecture, which was announced shortly before the launch of Intel’s Panther Lake Core Ultra 3 processors. This architecture is a further development of the existing Xe3 generation and is designed with an emphasis on energy efficiency and cost efficiency.
High-bandwidth memory scarce and too expensive
From a technical standpoint, the choice of LPDDR5X memory is noteworthy. Crescent Island has 160 gigabytes of memory. It uses LPDDR5X instead of high-bandwidth memory (HBM), which is common in Nvidia and AMD GPUs. HBM has become scarce and expensive due to rising demand in the AI sector. Intel’s choice of LPDDR5X therefore seems like a practical way to reduce costs and simplify production, though it may come at the expense of peak bandwidth.
The new GPU must be suitable for a variety of data types. This makes the product interesting for applications such as “tokens-as-a-service” and generative AI. Intel is also working on an open software stack compatible with various hardware types to offer developers greater flexibility. This software is currently being tested on Arc Pro B-series GPUs to enable early-stage optimizations.
Intel says it is aiming for an annual release cycle for new GPU generations. This pace is comparable to competitors’ product planning. The first Crescent Island units are expected to be delivered to customers in the second half of 2026.