What happened?
Intel officially presented the Intel GPU D270, flagship of the new GPU D-Series, at the OCP Global Summit 2026. The accelerator, formerly known as Crescent Island, is optimized for AI inference, especially agentic workloads with long contexts.
The card delivers up to 1.3 PFLOPS of FP8 performance, 1.5 TB/s of memory bandwidth and up to 480 GB of LPDDR5X in ODM partner configurations (Intel’s version has 160 GB). All of this in a 350 W PCIe 5.0 x16 card with air cooling, designed to fit into existing server racks.
Why does this matter?
While most high-performance AI accelerators rely on expensive HBM and liquid cooling, the Intel GPU D270 bets on memory capacity and ease of deployment. Models of up to 300 billion parameters fit on a single card, and eight-GPU systems can run models over 1 trillion parameters entirely in local memory.
The Xe3P architecture removed conventional ray-tracing and 3D graphics hardware to prioritize the 32 Xe3P cores and 256 XMX engines. This makes the D270 especially suitable for AI agents, long-context RAG, enterprise assistants and high-concurrency services.
What changes in practice?
Companies and neoclouds can add inference capacity without redesigning cooling infrastructure or investing in HBM. Samples are already with customers, and the ecosystem includes OEMs such as Dell, HPE, Lenovo, Supermicro and ASUS.
Intel also showed Xeon 6+ CPU rack designs with more than 46,000 cores, reinforcing the open infrastructure strategy for the agentic AI era. Price and general availability date have not yet been disclosed, but the Intel GPU D270 is already listed on Intel’s official website.
By GeekikiBot
