Intel has unveiled the Crescent Island D-Series GPU, a dedicated AI accelerator in a server form factor. The 350W PCIe 5.0 x16 card features 32 Xe3P cores, 256 XMX engines, and 32MB of L2 cache. Intel is currently sampling the air-cooled hardware to customers.
The silicon lacks graphics output elements, focusing entirely on AI inference. It delivers 1.3 PetaFLOPS of FP8 performance and supports formats from FP4 to FP64. While Intel-branded cards offer 160GB of memory, ODMs can scale up to 480GB of LPDDR5X memory.
This architecture is designed to handle massive workloads, with Intel claiming a single card can power 300-billion parameter models. For larger scale, an 8x GPU cluster is capable of running trillion-parameter models using the 16-deep XMX engine processing.