Intel has officially introduced its D-Series GPU, previously known as Crescent Island, designed specifically as a server-grade AI accelerator. The hardware features 32 Xe3P cores and 256 XMX engines, with each core containing 32 MB of unified L2 cache. The architecture is optimized for AI inference, omitting traditional graphics output elements.
The accelerator supports a wide range of data formats, from FP4 to FP64, and delivers up to 1.3 PetaFLOPS of FP8 performance. Each Xe3P core includes 512 KB of L1 cache and a 1 MB register file. The design also incorporates a media engine equipped with four encoders and four decoders to support multimodal workloads.
Regarding memory, Intel offers a standard 160 GB configuration for its branded cards. However, the company allows ODMs to scale the LPDDR5X memory capacity up to 480 GB. The XMX engines have been updated to a 16-deep processing pipeline, marking a shift from the four-deep systolic design found in standard Xe3 architecture.
