Overview
- Intel announced three complementary designs for agentic AI: Diamond Rapids as an enterprise Xeon-class processor, Crescent Island as a datacenter inference GPU, and Wildcat Lake as a client and edge SoC.
- Crescent Island is described as an Xe3P-based inference GPU with up to 32 Xe3P cores, 256 XMX engines, a 350 W air‑cooled PCIe reference card, and support for as much as 480 GB of LPDDR5X memory to boost model capacity at lower power.
- Diamond Rapids targets orchestration and high I/O with up to 256 cores, a large unified cache, 16 memory channels, PCIe Gen6 and CXL 3.0 support, and packaging built on Intel 18A and Foveros Direct 3D.
- Intel frames the portfolio around open packaging and chiplet standards such as UCIe to enable modular scaling, but the company’s claims hinge on software support, KV‑cache handling and partner integrations that are not yet independently validated.
- The market will judge the designs by real‑world metrics such as customer samplings, independent benchmarks and cost‑per‑token comparisons versus HBM4 rack solutions from rivals, with near‑term adoption depending on software, shipping timelines and total inference economics.