Intel names Crescent Island as GPU D-Series, adds D270 inference card
At OCP 2026, Intel officially named its Crescent Island data center inference accelerator the Intel GPU D-Series, with the first product being the GPU D270. Built on the Xe3P architecture with 32 Xe3P cores and 256 XMX engines, the card runs at 350W in a standard air-cooled PCIe Gen5 form factor and is rated at up to 1.3 PFLOPs of FP8 peak compute with 1.5 TB/s of LPDDR5X memory bandwidth (160GB standard, up to 480GB from partners). Intel says a single card can run models of roughly 300 billion parameters and an eight-card system can support models over 1 trillion parameters, targeting long-context, multi-tenant inference and latency-sensitive workloads such as robotics; sampling with server partners is underway, but pricing and broad launch timing were not disclosed.
As inference and agentic workloads dominate data center costs, a 350W air-cooled card targets enterprises that cannot or will not rebuild racks around liquid cooling.