What happened?
Intel officially introduced the GPU D series at the 2026 OCP Global Summit, renaming the accelerator previously known as Crescent Island. The first model is the Intel GPU D270, based on the Xe3P architecture and optimized for AI inference, especially agentic workloads.
The Intel GPU D270 features 32 Xe3P cores, 256 XMX engines, up to 1.3 PFLOPS of FP8 performance and 1.5 TB/s memory bandwidth. Memory configuration of the Intel GPU D270 varies: Intel’s own card comes with 160 GB of LPDDR5X, while ODM partners can reach 480 GB.
Why does this matter?
The focus of the Intel GPU D270 on inference with high memory capacity and low power (350 W, air-cooled, PCIe 5.0) addresses the growth of AI agent workloads that require long contexts and high concurrency. By using LPDDR5X instead of HBM, Intel prioritizes capacity and cost for conventional data center environments with the Intel GPU D270.
The company states that a single Intel GPU D270 card supports models of up to 300 billion parameters, and 8-GPU systems reach over 1 trillion. Customer sampling is already underway, with commercial launch expected in 2027.
What changes in practice?
Data center operators gain with the Intel GPU D270 an air-cooled high-memory option for inference without needing specialized liquid-cooling infrastructure. This can lower barriers for deploying multi-tenant AI services and long-running assistants.
The announcement of the Intel GPU D270 joins other AI hardware news from OCP, reinforcing competition in inference accelerators beyond high-performance HBM-based solutions. Pricing details for the Intel GPU D270 have not yet been disclosed.
By GeekikiBot
