Intel's new AI GPU packs 480 GB of RAM while consumers struggle to find memory

Intel just announced its GPU D-Series — formerly codenamed Crescent Island — an AI-focused chip stuffed with up to 480 GB of LPDDR5X memory. That's a lot of RAM to absorb when consumer DRAM is already in short supply and getting shorter.

Built on Intel's Xe3P architecture, the GPU D-Series delivers 1.5 TB/s of memory bandwidth and 1.3 petaflops of FP8 compute, all within a 350-watt power envelope. Intel is pitching it as infrastructure that simplifies deployment in existing data centers while supporting portability across Intel GPU platforms. The target workloads range from multi-tenant AI services to robotics and latency-sensitive inference tasks. Intel VP of AI Products Anil Nanduri put it plainly: "Inference is becoming the defining AI workload, and enterprises need infrastructure that delivers the right performance at the right cost."

For context, Intel's previous Gaudi 3 chip topped out at 128 GB of unified HBM — so the D-Series is a massive jump in raw capacity, even accounting for the switch from HBM to the slower but more power-efficient LPDDR5X. Intel's earlier Gaudi line never seriously threatened Nvidia in the AI accelerator market, and the follow-up Falcon Shores project was cancelled outright. The D-Series pivots away from training workloads entirely, focusing on inference instead — a smarter lane to compete in given Nvidia's stranglehold on the training side.

For PC gamers watching memory prices tick upward, this is another reminder of where production capacity is being directed. Even future consumer wins — like a rumored memory bump in a next-gen RTX card — will likely arrive with AI-optimized specs and AI-adjusted pricing to match.