Intel Crescent Island GPUs Pack Up To 32 Xe3P Cores, Optimized For Agentic AI With Low-Cost LPDDR5X That Reaches Up To 480 GB Capacity

Key Points:

• Intel has unveiled its Crescent Island GPU, an AI inference accelerator featuring 32 Xe3P cores and up to 480 GB of power-efficient LPDDR5X memory, designed for air-cooled data center and workstation environments with a 350W TDP.
• Based on the new Xe3P architecture, Crescent Island emphasizes high memory capacity, sustained compute performance, and broad AI model support, including language, reasoning, multi-modal, and diffusion models, optimized for agentic AI workloads.
• The GPU architecture integrates 8 Vector Engines and 8 Xe Matrix Extensions per core, with a total of 256 Vector Engines and 256 XMX units across 32 cores, supported by a large cache hierarchy and flexible data format compatibility from FP4 to FP64.
• Intel leverages LPDDR5X memory to reduce power consumption and costs compared to competitors using HBM memory, targeting improved performance-per-watt and total cost of ownership, with partner options supporting memory capacity up to 480 GB.
• Crescent Island supports an open software ecosystem compatible with frameworks like vLLM and NVIDIA Dynamo, enabling heterogeneous agent orchestration, with customer sampling planned for the second half of 2026.