Intel Outlines Crescent Island GPU for AI Inference; Sampling Expected in 2H 2026

Intel is preparing Crescent Island, an inference-focused data center GPU designed for enterprise systems, with customer sampling expected in the second half of 2026.

Intel introduced the GPU at the 2025 OCP Global Summit. The company describes it as a 350-watt, air-cooled PCIe card based on its Xe3P architecture, with 32 Xe cores and 256 XMX engines.

The reference design includes 160 GB of LPDDR5X memory, while the broader platform supports up to 480 GB.

Crescent Island supports data types ranging from FP4 to FP64 and is intended for agentic artificial intelligence and other inference workloads. The design can operate alongside Xeon processors and Intel’s open AI software stack.

Intel is emphasizing memory capacity, power efficiency and deployment in conventional air-cooled servers. Intel has not published comparable bandwidth figures or independent performance results for the GPU.

“AI is shifting from static training to real-time, everywhere inference—driven by agentic AI,” Intel Chief Technology Officer Sachin Katti said.

“Scaling these complex workloads requires heterogeneous systems that match the right silicon to the right task, powered by an open software stack,” Katti said.

Crescent Island forms part of Intel’s effort to compete in AI infrastructure. Intel has not announced that customer sampling has started.

Customer sampling is expected in the second half of 2026.