Today, Intel introduced its upcoming Crescent Island data center GPU designed specifically for AI inference workloads and built on the company's next-generation Xe3P "Celestial" architecture. The new GPU will focus on performance-per-watt efficiency, cost optimization and air-cooled operation, making it Intel's latest push into the growing AI accelerator market. Intel didn't release many technical details, so far announcing that Crescent Island will feature 160 GB of onboard LPDDR5X memory, an unusually large configuration for a GPU; however, this will help in handling large language models (LLMs) and other memory-intensive AI workloads. The use of LPDDR5X suggests a wide memory interface potentially up to 640 bits depending on whether Intel opts for a single high-end GPU die or a dual-GPU configuration. This setup provides high bandwidth and energy efficiency, although LPDDR5X's dual 16-bit channel design rules out butterfly connection modes common to GDDR6/7 memory.

Intel says customer sampling for Crescent Island will begin in the second half of 2026, with broader availability expected in 2027. The company is currently refining its open-source software stack using existing Arc Pro B-Series GPUs to ensure smooth scaling to Xe3P hardware. While detailed specifications remain under wraps, Crescent Island will compete with next-generation AI accelerators such as AMD's Instinct MI450 series and NVIDIA's Vera Rubin.