Intel Details Crescent Island GPU with Xe3P Architecture for AI Inference
Intel detailed its new Crescent Island GPU at the Hot Chips conference, featuring the Xe3P architecture designed specifically for AI inference and agentic AI workloads. The GPU supports up to 480GB of LPDDR5X memory in partner versions and operates at a 350W TDP. By removing 3D and ray tracing hardware, Intel is optimizing silicon space purely for AI compute, reflecting an industry-wide shift toward specialized AI hardware. The massive memory capacity and deeper systolic array design will significantly boost performance for large language models and complex AI agents. The Xe3P architecture expands the XMX (Xe Matrix Extensions) units to a 16-deep systolic design (up from 4-deep in Xe2/Xe3) and adds native support for FP8 and FP4 data formats. The PCIe Gen5 x16 card features 32 Xe cores, providing a total of 256 Vector Engines and 256 XMX Engines.
## BACKGROUND
Intel's Xe architecture is a scalable GPU architecture used across integrated graphics, discrete GPUs, and data centers. XMX (Xe Matrix Extensions) are specialized hardware blocks within Intel GPUs designed to accelerate matrix multiplication, which is the foundational mathematical operation for AI and deep learning.