Crescent Island GPU

Overview

Crescent Island is Intel's next-generation inference-optimized GPU, announced by Intel executive Sachin Katti at the OCP Global Summit 2025 keynote. Built on Intel's Xe 3P IP architecture, it offers up to 160 GB of LPDDR5x memory with 1.5 TB/s memory bandwidth—deliberately optimized for inference memory capacity rather than training throughput. It supports data types from MXFP4 through FP32 and FP64, and is designed in a PCIe card form factor to deliver low-cost, power-optimized inference deployable in enterprise servers without specialized rack infrastructure. Samples are targeted for the second half of 2026.

Intel positions Crescent Island as filling a critical gap in the AI accelerator market for cost-effective inference at the PCIe server tier, serving as the "Scale" phase in Intel's three-phase annual GPU cadence (Deliver → Scale → Expand), with the "Expand" phase to introduce the next-gen Shores product line.

Sign in to read the full article.

Sign In