Particle.news

Download on the App Store

Intel Introduces Crescent Island, an Xe3P Inference GPU With 160GB LPDDR5X

Intel positions the GPU for air‑cooled inference in data centers, emphasizing LPDDR5X capacity.

Overview

  • Intel announced the data center accelerator at the OCP Global Summit and set customer sampling for the second half of 2026.
  • The design targets cost‑ and power‑sensitive deployments and is tuned for real‑time inference and token‑based services.
  • Crescent Island is built on the Xe3P microarchitecture, which Intel says is optimized for performance per watt.
  • The card integrates 160GB of LPDDR5X, a choice outlets report favors cost and availability over HBM but carries LPDDR5X channel constraints.
  • Intel is developing a unified, open software stack for heterogeneous AI systems, testing it now on Arc Pro B‑Series GPUs, and has not provided performance estimates.