Intel's AI accelerator finally has a name, a spec sheet, and a shipping form factor.
At OCP 2026, Intel rebranded its Crescent Island chip as the GPU D-Series and disclosed real numbers for the first time: 1.3 petaflops of FP8 compute and 1.5 TB/s of memory bandwidth. The chip is built on Intel's Xe3P architecture, ships in an air-cooled PCIe card capped at 350W, and carries up to 480GB of LPDDR5X memory. Intel says servers can scale up to eight GPU D-Series cards, similar to Nvidia's HGX layout, enough to run inference on models beyond a trillion parameters. Customer sampling is underway, though Intel hasn't set a date for broad availability.
The numbers only make sense next to Nvidia's Vera Rubin NVL72, where each of 72 liquid-cooled Rubin GPUs claims 17.5 petaflops of FP8 - more than 13 times the GPU D-Series's throughput. Intel isn't trying to out-muscle that rack; it's betting that plenty of inference workloads don't need liquid cooling or kilowatt power budgets, just a chip that slots into a standard air-cooled server.
Intel has made similar bets before, with the Data Center GPU Max line that never found a market. The real test for GPU D-Series is whether its software stack and that broad ecosystem of partners actually show up once the chips ship in volume.