Intel's Crescent Island becomes GPU D-Series AI inference accelerator
Intel used OCP 2026 to rename its Crescent Island accelerator as the GPU D-Series. The chip is rated for 1.3 PFLOPS of FP8 compute, 1.5 TB/s of memory bandwidth, up to 480GB of LPDDR5X, and a 350W air-cooled PCIe design. Intel says the product is aimed at inference workloads, including large language models and multi-tenant AI services, and can scale to eight GPUs.
Intel disclosed the GPU D-Series at OCP 2026, giving its Crescent Island accelerator a formal product name. The part is built on the Xe3P architecture and is intended for inference rather than graphics, with a 350W air-cooled PCIe card design, up to 480GB of LPDDR5X memory, 1.3 PFLOPS FP8 peak compute, and 1.5 TB/s bandwidth.
Intel says up to eight cards can work together, enough for models exceeding one trillion parameters. Customer samples are underway, though launch timing remains unspecified. The company also points to its shared Xe software base and ecosystem partners. By contrast, Nvidia's Rubin chips target far higher FP8 throughput with liquid cooling.
If adopted, air-cooled inference accelerators could lower barriers for smaller cloud providers and enterprises, potentially expanding access to AI services. Data center operators may benefit from standard rack integration, while energy use and hardware demand could still shape local infrastructure. Researchers, developers, and businesses relying on shared AI services could see more options and competition, though actual effects depend on software maturity, pricing, and availability.