Somewhere a PC builder is pricing out a 32 GB DDR5 kit and wincing. Meanwhile Intel has just announced a single graphics card that can carry 480 GB of the same family of memory.
Intel officially named its Crescent Island accelerator the GPU D-Series this week at OCP. The D270 is the first product in the family. It is built on the Xe3P architecture and pairs up to 480 GB of LPDDR5X with 1.5 TB/s of memory bandwidth. Intel rates it at 1.3 petaflops of FP8 compute within a 350W power envelope. It’s a PCIe card designed for air-cooled servers and workstations. Intel’s own reference configuration carries 160 GB, and partners can scale to the full 480 GB.
Intel’s VP of AI Products and GTM, Anil Nanduri, framed the goal as making high-capacity inference practical and economical to scale. Intel lists target uses from multi-tenant AI services to robotics and other latency-sensitive workloads.
Gaudi 3 carried 128 GB of HBM and never seriously threatened NVIDIA. Later that year, Intel itself conceded it couldn’t compete with NVIDIA at the high end of AI. The D-Series reads like a company that took that lesson to heart. Rather than chase NVIDIA on HBM bandwidth, Intel is betting on capacity per dollar and per watt. That suits inference, where long context windows and KV caches eat memory faster than they eat compute.
LPDDR5X is not HBM. It comes off the same kind of production lines that feed laptops, handhelds, and phones, so every 480 GB card is mobile DRAM that isn’t going into consumer devices. With memory supply already tight across the board, another AI product built on commodity DRAM is not what the RAM shortage needed.
For gamers, there is one silver lining. Xe3P is also the architecture expected to power Intel’s next-generation Arc client GPUs. Whatever efficiency work Intel did for a 350W inference card should carry over to the desktop part.
Intel hasn’t announced pricing or partner systems yet. Given the memory market, we wouldn’t hold our breath for “cheap.”
Discussion
Join Discussion →No replies yet. Be the first to reply on the forum.