Intel used Hot Chips 2026 to lay out its agentic AI server roadmap, pairing a 256-core Xeon 7 processor built on its 18A-P process with an air-cooled inference GPU that swaps HBM for cheaper LPDDR5X memory.
Diamond Rapids, the next Xeon 7, packs 256 Panther Cove-X P-cores, a 50 percent jump over Granite Rapids that matches AMD’s top EPYC Venice. It is Intel’s first CPU on 18A-P and introduces a fan-out fabric architecture with a unified memory fabric, 1.28 GB of cache, and memory speeds up to 12,800 MT/s. Compute, memory, and I/O are disaggregated into tiles, including a new core building block.
On the accelerator side, Crescent Island carries 32 Xe3P cores with up to 480 GB of LPDDR5X, targeting air-cooled data centers and workstations for inference and agent workloads. Intel argues the memory choice cuts cost sharply against the HBM3E and HBM4 stacks used by Nvidia’s Rubin and AMD’s MI450X, and that FP4 and FP8 support suits token-as-a-service economics.
Both designs lean entirely on Intel silicon, an intentional bet on 18A-P after Panther Lake and Clearwater Forest established the node family. Intel is also expanding its Israeli R&D role as the server push accelerates, according to Calcalist.
