Intel unveils Crescent Island GPU with Xe3P architecture for AI inference

ByEnigma 808

Intel has revealed details about its upcoming data center GPU, Crescent Island, which introduces the new Xe3P architecture. Unveiled at the OCP Global Summit, Crescent Island will be the first to use Xe3P, later appearing in Intel’s Celestial and Nova Lake-S products. Designed for AI inference workloads, it promises better performance and power efficiency compared to the current Xe3 architecture in Panther Lake.

A key feature of Crescent Island is its memory setup, supporting up to 160GB of LPDDR5x memory instead of the usual GDDR6 or GDDR7. While LPDDR5x is slower in bandwidth, it is more efficient and compact, making it ideal for servers with power and thermal constraints. The GPU is built for flexibility, supporting multiple data formats like FP8, INT8, and BF16, making it suitable for AI tasks.

Crescent Island focuses on inference rather than full AI training, aiming to complement Intel’s broader data center ecosystem. Intel plans to ship samples in the second half of 2026, with a full launch expected in 2027.

Author

  • Enigma XO avatar (80x80)

    Besides writing and being a content creator, Enigma 808 sometimes likes to watch paint dry on walls, listening to ASMR while timing snails racing. Such is life.