Unlike traditional GPUs focused on game rendering, this series of accelerators is specifically designed to handle long-context processing and concurrent multi-path inference, marking another strong push by Intel into the AI inference space.Author and source: AIBase
At the Open Compute Project (OCP) Global Summit, Intel officially unveiled its new GPU D series (codenamed Crescent Island, officially listed as D270), designed specifically for AI inference. Unlike traditional GPUs focused on gaming rendering, this series of accelerator cards is engineered to address long-context processing and concurrent multi-path inference, marking another strong push by Intel into the AI inference market.
Key Highlights: High VRAM, Low Power Consumption, and Seamless Compatibility
Massive VRAM limit: 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 interconnect system can smoothly run ultra-large models exceeding 1 trillion parameters. (Note: The standard Intel OEM PCIe card comes with 160GB of VRAM; the 480GB version is a custom limit available to ODM partners.)
Computing and Energy Efficiency Performance: Built on the all-new Xe3P architecture, the GPU offers a memory bandwidth of 1.5 TB/s, delivering 1.3 PFLOPS of computing power at FP8 precision, with a total card power consumption capped at 350W.
Out-of-the-box deployment: Features a standard PCIe form factor and air-cooling design, eliminating the need for costly liquid cooling upgrades to existing data centers, significantly lowering the barrier to enterprise deployment.
Why choose LPDDR5X?
Typically, high-power top-tier AI accelerator cards prioritize high-bandwidth HBM memory, while LPDDR5X is commonly found in power-sensitive mobile devices. Intel’s selection of LPDDR5X for its data center-grade GPU is primarily driven by its exceptional energy efficiency and massive capacity.
Directly integrating up to 480GB of capacity onto the accelerator card enables large model weights and context data during inference to remain largely resident in the card’s memory. This significantly reduces network communication overhead caused by splitting models across multiple cards, making it ideally suited for high-frequency use cases such as long prompt processing, enterprise-grade AI services, multi-tenant concurrency, and real-time robotic responses.
Listing progress
Currently, the Intel GPU D series has entered the customer sampling phase, and the public price and official release date have not yet been announced. With the launch of this significant product, competition in the data center AI inference market is undoubtedly set to enter a new phase of escalation.