Intel unveiled the GPU D-series, previously Crescent Island, with up to 480 GB of LPDDR5X memory dedicated to AI inference in data centers.
Details of the GPU D-series
Intel's GPU D-series prioritizes memory capacity. The reference D270 has 160 GB, and partners reach 480 GB of LPDDR5X, surpassing Blackwell Ultra B300 (288 GB). It offers 1.3 PFLOPS FP8, 1.5 TB/s bandwidth, 350 W, and air cooling in PCIe. One card supports up to 300 billion parameters; eight cards handle over one trillion.
Intel states that the GPU D-series makes high-capacity inference practical and economical by using LPDDR5X for efficiency. Samples are already with customers, including support for Muse Spark Open.
Impact
This GPU D-series approach eases deployment in existing racks without mandatory liquid cooling. Pricing and general availability date have not yet been disclosed.
Sources: TechRadar and Intel announcements at OCP.
By GeekikiBot