Intel's newest AI chip holds 480 GB of memory, more than 60 times what ships in a typical gaming PC.
Intel unveiled the GPU D-Series, codenamed Crescent Island, built on its Xe3P architecture. It packs up to 480 GB of LPDDR5X memory, 1.5 TB/s of memory bandwidth, and 1.3 petaflops of FP8 compute, all inside a 350-watt power envelope. Intel says the chip targets AI inference rather than training, aiming at jobs like multi-tenant AI services and robotics. It follows the failed Gaudi lineup and the cancelled Falcon Shores GPU, Intel's earlier attempts to compete with Nvidia in AI hardware.
The real story here is the memory, not the compute. LPDDR5X is the same memory family used in phones and laptops, and it's cheaper than the HBM Nvidia and AMD lean on, which means Intel is pulling huge volumes of it at a moment when PC and laptop RAM is already in short supply. Any extra memory headed for gaming GPUs like the RTX 5080 next year will be competing with chips built for inference farms, not against them.
Intel finally found a market that wants more memory than anyone really needs. It just isn't the market with a graphics card in it.