Intel will offer up to 480 GB LPDDR5X memory option in its new GPU codenamed Crescent Island, prepared for the data center and workstation side. The new GPU, built on the The card is built on the Xe3P microarchitecture. Xe3P is positioned as a more advanced version of Intel’s Xe3 architecture and offers a structure shaped by performance/watt efficiency on the Crescent Island side. The most striking aspect of the new GPU was its memory capacity.
Intel’s reference PCIe card comes with 160 GB LPDDR5X memory. However, card partners will be able to increase this capacity up to 480 GB in special ODM branded designs. This capacity is above the 288 GB HBM4 level on the NVIDIA Vera Rubin side and the 432 GB HBM4 level on the AMD Instinct MI450X side. This comparison is made only on the amount of memory; Bandwidth, computing power and system scaling are considered as separate topics.
Crescent Island prefers LPDDR5X memory instead of HBM3E or HBM4 used in classic high-end data center GPUs. The use of LPDDR5X allows the card to have a more controlled structure in terms of power and cost. According to the information shared by Intel, the card will operate at 350W power level in air-cooled PCIe design. This structure is prepared for air-cooled data center servers and workstations, unlike liquid-cooled rack systems.
Intel states that the LPDDR5X memory structure allows dense channel design and provides an increase in bandwidth. The board’s strategy on the memory side opens a different path for AI inference systems that require high capacity at a time when HBM supply is more expensive and limited. At this point, Crescent Island is shaped around high memory capacity, power efficiency and total cost of ownership rather than raw bandwidth race.
Crescent Island is a separate product from the Arc graphics card family on the consumer side. The card was developed not as a gaming-focused graphics card, but as an accelerator primed for data center AI inference. The shared technical information includes the information that the product is designed for GPGPU and AI compute side rather than traditional graphics and 3D workloads. The new GPU comes with wide data type support for agentic AI systems, AI inference services and tokens-as-a-service providers.
Data formats ranging from FP4 to FP64 are supported on the card. This range provides a wide range of uses, from low-precision AI inference workloads to scenarios requiring more precise calculations. Intel is also preparing an open and unified software stack for Crescent Island. The company is testing the software layer it developed for heterogeneous AI systems on existing Arc Pro B series GPUs. This software side paves the way for installations where different accelerators run in the same system via Intel Xeon processors and Intel GPUs.
Performance values for Crescent Island have not been shared yet. Intel will begin customer sampling of the card in the second half of 2026. In this process, in addition to the reference 160 GB design, card partners are expected to prepare special versions with LPDDR5X memory that can go up to 480 GB.
Comments
You can write your views about this story. Comments may be moderated according to site settings.