At the OCP Global Summit, Intel officially unveiled the mysteries of the new GPU D series for AI inference (codenamed Crescent Island, marked as D270 on the official product page). Unlike traditional GPUs focused on game rendering, these acceleration cards are designed specifically to handle long context processing and multi-concurrent inference, marking another strong push from Intel in the AI inference field.
Key standout features: high memory capacity, low power consumption, and seamless compatibility
Maximum VRAM Capacity: The D series supports up to 480GB of LPDDR5X VRAM. Official data shows that a single card can handle models with up to 300 billion parameters, while an eight-card interaction system can smoothly run ultra-large models with over 1 trillion parameters. (Note: The standard version of Intel’s original PCIe cards has 160GB of VRAM; the 480GB version is a customized limit available for ODM partners.)
Computing and energy efficiency performance: Based on the new Xe3P architecture, the memory bandwidth reaches 1.5TB/s, and under FP8 precision, it can output 1.3PFLOPS of computing power. Meanwhile, the total power consumption of the entire chip is limited to 350W.
Out-of-the-box deployment: Utilizing standard PCIe form factor and air cooling design, no expensive liquid cooling transformation of the existing data center is required, significantly reducing the deployment barrier for enterprises.
Why choose LPDDR5X?
Generally, high-power top-tier AI acceleration cards prefer high-bandwidth HBM memory, while LPDDR5X is commonly used in mobile devices that are sensitive to power consumption. Intel’s choice of LPDDR5X for data center-level GPUs is primarily due to its extremely high energy efficiency ratio and large capacity.
A capacity of up to 480GB directly mounted on the accelerator card allows the weight parameters and context data during inference for large models to remain as much “in memory” as possible on a single card. This significantly reduces the network communication overhead due to the division of the model across cards, making it highly suitable for high-frequency use cases such as long Prompt processing, enterprise-level AI services, multi-tenant concurrency, and real-time robot responses.
Launch Progress
Currently, the Intel GPU D series has entered the sample distribution phase for customers, and there is no public pricing or exact launch date yet. With the release of this important product, competition in the data center AI inference market will undoubtedly undergo a new round of intensification.