Intel GPU D270 bets on laptop memory for air-cooled inference

Intel’s GPU D270 uses LPDDR5X memory instead of High Bandwidth Memory, targeting enterprise inference deployments in standard air-cooled racks. Formerly known as Crescent Island, the card combines 480 GB of memory with a 350-watt power draw. Intel’s premise is that model capacity, serving costs and power efficiency can matter more than peak memory bandwidth, potentially avoiding liquid-cooling upgrades and reducing reliance on concentrated HBM supplies.

The strategy accompanies strong growth in Intel’s Data Center and AI segment, which reported $7.318 billion in revenue in Q3 2026, up 78% year-over-year. However, Xeon server CPUs remain the dominant revenue driver. The GPU D270 is still in customer sampling, has not reached volume production and has no separately disclosed revenue contribution. Those earnings therefore do not establish commercial success for the accelerator, and independent production benchmarks have not yet validated its inference economics.

Intel confirmed compatibility with NVIDIA’s Dynamo inference orchestration framework, potentially easing integration into existing serving environments. A multiyear Oracle collaboration also supports Intel’s broader strategy, with future cloud infrastructure planned around Diamond Rapids Xeon processors. Supply constraints continue to limit Intel’s ability to meet demand, while U.S. export-license requirements restrict sales of the GPU D-series and certain Xeon configurations to Chinese cloud customers.