Intel unveils GPU D-series for AI inference

Intel has revealed technical specifications for its GPU D-series, a new line of data center GPUs designed specifically for AI inference tasks. Announced at the OCP Global Summit, the series—internally known as ‘Crescent Island’—utilizes the Xe3P architecture to support large-scale AI models.

The GPUs feature up to 480GB of LPDDR5X memory and a memory bandwidth of 1.5TB/s, delivering 1.3 PFLOPS of FP8 computing performance within a 350W power envelope. A single card can handle models with up to 300 billion parameters, while systems utilizing eight GPUs can support models exceeding 1 trillion parameters. To ensure ease of integration, the cards use a standard PCIe form factor and air-cooling, allowing deployment in existing data center environments without specialized cooling infrastructure.

In addition to the GPUs, Intel showcased new rack configurations powered by Xeon processors developed with partners. These designs follow the Open Rack v3 (ORv3) standard, with some Xeon 6+ configurations capable of supporting over 46,000 CPU cores per rack. Intel is also promoting an open software ecosystem to ensure compatibility with existing Xe software and new AI models.