Intel reveals Crescent Island inference GPU design

Intel has published a bare-die image of its Crescent Island GPU, offering a closer look at the Xe3P architecture behind its recently announced D-series inference cards. The highest-end D270 features 32 Xe3P cores, 256 XMX engines, and 32 MB of unified L2 cache. The card targets AI inference rather than training and supports up to 480 GB of LPDDR5X.

The XMX engines move from the four-deep systolic design used in regular Xe3 to 16-deep processing. Xe3P also supports FP4 alongside FP8, with register files of 1 MB per core and an L1 cache of 512 KB per core. The GPU includes 256 Vector engines, with Vector and XMX engines each arranged in groups of eight per Xe3P core.

Designed for long-context and agentic workloads, Xe3P omits the graphics processing elements needed to output imagery. A media engine with four encoders and four decoders could support multimodal tasks involving video generation and understanding.