At the OCP Global Summit 2026, Intel gave its Crescent Island AI accelerator an official product name and disclosed final performance figures. The chip will be sold as the Intel GPU D-Series, with the first card named GPU D270. This is Intel's first commercial data center GPU since the Ponte Vecchio generation, following several canceled successors.
Both Tom's Hardware and ServeTheHome report the same headline specs: 1.3 PFLOPS of FP8 matrix throughput, 1.5 TB/s of memory bandwidth, and an air-cooled PCIe CEM form factor with a 350W TDP. The sources diverge on memory configuration, however. Tom's Hardware says the product line supports up to 480GB of LPDDR5X, while ServeTheHome adds that Intel's own GPU D270 card ships with 160GB; the 480GB version would come only from ODM partners using 24GB LPDDR5X chips.
Intel is positioning the D-Series for inference workloads such as large language models, long-context inference, and multi-tenant services. Intel also claims that eight D-Series GPUs in a standard air-cooled server can handle models with more than one trillion parameters. ServeTheHome notes that the 350W card is much less power-hungry than AMD's flagship PCIe accelerator, the 600W MI350P, but its theoretical FP8 peak is roughly 40% lower; how that translates into real-world efficiency remains to be seen.
Tom's Hardware frames the announcement against Nvidia's liquid-cooled Rubin systems, emphasizing that the GPU D-Series targets a different performance and infrastructure envelope. Both outlets note that customer sampling is underway and that Intel has not announced a broad launch. Whether the GPU D-Series gives Intel a needed AI accelerator foothold after years of setbacks is still an open question.