Intel has introduced its new AI inference GPU for data centers, called Crescent Island, which uses the Xe3P architecture. The Xe3P graphics architecture will power Blue Team’s next-generation Crescent Island GPU for workstation AI inference, which is scheduled to arrive in the second half of 2026.
After not fully reaching its AI goals in the data center market, Intel is now slightly changing its strategy. The company appears to be focusing on efficiency and larger memory capacity with a completely new product called Crescent Island. This is the same GPU that was previously teased during the Intel Tech Tour 2025. The Intel Crescent Island GPU is built on the new Xe3P architecture, which is the same graphics technology the company showcased last week during its Panther Lake and Xe3 presentations.
The New Architecture Will Be An Improvement Over The Xe3 Design!
The new architecture will be an improvement over the Xe3 design, and for consumer devices it will appear in the next-generation Arc C-Series graphics family. However, Xe3P will be even more scalable, as it can be used in everything from integrated GPUs in client devices to AI GPUs in data centers. Intel Crescent Island will be optimized for both power efficiency and cost. It is designed for air-cooled data center setups and will focus on handling AI inference tasks.
According to Intel, the Xe3P graphics architecture used in Crescent Island will be optimized for performance per watt. The GPU will include a large 160 GB memory capacity using the LPDDR5X standard. Interestingly, Intel has chosen to use LP5X for this model.
Competitors Provide High-End HBM Memory
Competitors like NVIDIA and AMD are providing their data center AI products with high-end HBM memory, such as HBM3E, and they are already discussing HBM4 for their next-generation chips like Rubin and MI400.
However, at the same time, getting HBM has become challenging because of high demand, which has also caused prices to rise. Using LPDDR5X memory could give Intel a major advantage in terms of cost and performance. In addition, the architecture will support a wide range of data types, which is well suited for “Tokens-as-a-service” providers and AI inference workloads. Intel is already testing its open and unified software stack on the Arc Pro B-series lineup, which means future versions will benefit from these optimizations early. The company plans to begin customer sampling of the Crescent Island GPU in the second half of 2026, so more information about the GPU will be revealed in the coming months.
