Intel maps agentic AI hardware from data centers to edge

Intel used Hot Chips 2026 to outline a broader hardware strategy for agentic AI, moving beyond a CPU-centered approach toward multiple architectures working across the computing stack. The plan centers on Xeon Scalable 7, codename Diamond Rapids, alongside Crescent Island, a data center GPU for inference, and Wildcat Lake, the architecture behind new Core Series 3 processors.

Diamond Rapids is positioned as the general-purpose foundation for enterprise-scale agentic AI, handling orchestration and processing between accelerator operations. Built with Intel’s 18A-P manufacturing technology, the design uses adaptable compute blocks, a unified memory fabric and flexible I/O. Intel said the architecture scales to 256 CPU cores, uses Efficiency cores, includes as much as 1.28GB of last-level cache, supports 16 memory channels at up to 12,800 MT/s, and offers 128 lanes of PCIe 6.0 and CXL 3.0 connectivity.

Crescent Island targets inference with 32 Xe3P-based Xe cores, 256 XMX engines and support for up to 480GB of LPDDR5X memory. Intel is aiming for a maximum TDP of 350 watts so the GPU can run in existing air-cooled data centers. Wildcat Lake brings AI features to lower-cost notebooks and edge devices with new CPU cores, integrated Xe3 graphics, XMX acceleration and an NPU capable of up to 17 TOPS.