Intel names Crescent Island D-Series GPU with up to 480 GB of LPDDR5X

Intel has introduced its ‘Crescent Island’ GPU under the D-Series name, positioning the server-form-factor card as a dedicated AI accelerator after earlier efforts stalled. The design includes 32 Xe3P cores, 256 XMX engines, 32 MB of unified L2 cache and up to 480 GB of LPDDR5X memory.

The D-Series GPU is rated for up to 1.3 PetaFLOPS of FP8 performance and supports data formats from FP4 to FP64. A media engine with four encoders and four decoders is built in, potentially helping multimodal workloads that depend on video generation and understanding.

Intel-branded ‘Crescent Island’ cards ship with a 160 GB memory configuration, while ODMs can equip the accelerator with up to 480 GB of LPDDR5X memory. Xe3P also changes the XMX pipeline from the four-deep systolic design used on regular Xe3 to 16-deep processing, with 1 MB register files per core and 512 KB of L1 cache per core.

The chip includes 256 Vector engines and 256 of the new XMX engines, organized in groups of eight per Xe3P core. Intel removed graphics processing elements for image output, reflecting a design focused on AI inference workloads requiring long context and agentic tasks.