Intel details its GPU D-series (Crescent Island) inference chip: up to 480GB memory in 350 watts

At the Open Compute Project Global Summit, Intel shared new technical details of its Intel GPU D-series, formerly code-named Crescent Island, an inference GPU built on its Xe3P architecture. Intel says it combines up to 480GB of LPDDR5X memory, 1.5TB/s of memory bandwidth and 1.3 petaflops of FP8 compute within a 350-watt power envelope. The 480GB figure applies to partner-built cards; Intel's own branded PCIe card has 160GB. Intel says the GPU supports models of up to 300 billion parameters on a single card and more than one trillion parameters in an eight-GPU system. It is an air-cooled, standard PCIe card meant to fit existing data centers without special rack or cooling setups, and Intel says customer sampling is underway. Intel also showed Xeon-based rack designs, some reaching more than 46,000 cores per rack at 133.45kW total rack power. All figures are Intel's own and have not been independently tested.