Intel Unveils Crescent Island: 350W Air-Cooled PCIe AI Inference Card with 32 Xe3P Cores, Up to 480GB

Intel unveiled details of its next-generation accelerator, codenamed Crescent Island, at the Hot Chips 2026 conference. The product targets PC, workstation, and data center edge deployments, with a focus on high-speed artificial intelligence inference. Built on Xe3P GPU IP, it takes the form of a 350W air-cooled PCIe expansion card.

Crescent Island is equipped with 32 Xe3P cores, 256 XMX matrix engines, 256 vector engines, and 32MB of unified L2 cache. Intel's own branded version will carry 160GB of LPDDR5X memory, while ODM partners are permitted to expand memory capacity to a maximum of 480GB of LPDDR5X to support local deployment of very large models.

On the computing architecture side, Xe3P upgrades the XMX engine processing pipeline, extending the four-deep systolic array design of the standard Xe3 to a 16-deep processing structure, and adds support for the FP4 data format while remaining compatible with FP8. Each Xe3P core also comes with a 1MB register file and 512KB of L1 cache to serve high-throughput AI computing.

Crescent Island is not a traditional graphics card intended for rendering or display output. The Xe3P chip omits the graphics processing unit used for image output, with its design focus shifted entirely toward AI inference workloads, including long-context processing and agent-based tasks. Intel said the product will primarily serve local PCs, workstations, and edge data centers that require rapid AI inference capabilities.

Taking trillion-parameter models such as Kimi K3 as an example, local deployment requires roughly 2TB of aggregate system memory. A workstation equipped with four Crescent Island accelerator cards, each carrying 480GB of memory, could in theory meet this capacity requirement.