THE CRUNCH
Intel has revealed its "Crescent Island" accelerator as a D-Series GPU, a dedicated AI inference card in a server form factor. The card carries 32 Xe3P cores, 256 XMX engines, 32 MB of unified L2 cache and up to 480 GB of LPDDR5X memory, with Intel rating it at up to 1.3 PetaFLOPS of FP8 performance across formats from FP4 to FP64.
The silicon is built purely for AI inference rather than graphics: there are no display outputs at all, and Intel positions it for workloads needing long context and agentic tasks. A media engine with four encoders and four decoders could help with multimodal work involving video generation and understanding. The Xe3P's XMX engines have also been redesigned, moving from a four-deep systolic layout on regular Xe3 to 16-deep processing.
Memory comes in two tiers. Intel-branded cards ship with 160 GB of LPDDR5X, while the company allows its ODMs to equip the card with up to 480 GB for maximum capacity. Intel suggests a single card can serve models up to 300 billion parameters, with an eight-GPU cluster handling trillion-parameter models.
Practically, the card is a 350 W, air-cooled PCIe 5.0 x16 unit, so it fits conventional servers without liquid cooling, and Intel leans on its common Xe software foundation and support for established AI frameworks to ease deployment. Customer sampling is underway, but pricing, availability and independent performance figures remain undisclosed.