Intel Crescent Island

Intel Crescent Island is an inference-optimised datacenter GPU built on the Xe3P microarchitecture, announced on 14 October 2025 and detailed at Computex 2026. Its defining decision is memory. Crescent Island uses LPDDR5X rather than HBM, trading bandwidth for capacity, cost and supply availability at a moment when HBM is committed years ahead. Intel's own reference card carries 160 GB, and Intel states the design delivers up to 480 GB of capacity, a ceiling intended for board partners building higher-capacity accelerators. Intel positions it as an air-cooled PCIe add-in card for standard enterprise servers rather than liquid-cooled racks, aimed at agentic AI and tokens-as-a-service inference. A board power of 350 W and a 480 GB capacity ceiling are widely reported and were previously recorded here, but neither appears in Intel's announcement or in either of Intel's Computex 2026 newsroom releases, which do not mention this GPU at all, so both are held out of the specification fields rather than published unverified. Intel describes support for a wide range of datatypes and microscaling formats, from native FP4 and MXFP4 through to FP64. Beyond that, Intel has published no throughput at all: no FLOPS, no TOPS, no memory bandwidth, no clock speeds, no Xe core count and no process node. Those fields are stored as NULL here rather than estimated. The bandwidth figures circulating in the press, roughly 684 GB/s and roughly 1.5 TB/s, are both third-party arithmetic on leaked PCB photographs, they assume different bus widths, and they disagree with each other by more than a factor of two. Customer sampling begins in the second half of 2026.