Loading...
Loading...
Intel is advancing its datacenter AI and server roadmap with two major reveals. Crescent Island, an upcoming Xe3P-based inference GPU due late 2026, targets token-efficiency and broad numeric formats while offering up to 480 GB of LPDDR5x memory on a 350 W PCIe AIC card and an open-source software stack. Looking further ahead to 2027, Intel’s Diamond Rapids Xeon family on an evolved 18A-P process promises ~50% more cores, higher single-thread performance, doubled memory bandwidth via more channels and faster memory, and PCIe Gen6—together signaling a push for greater compute density and I/O throughput in cloud and AI deployments.
Intel's Crescent Island GPU and Diamond Rapids Xeon CPUs reshape datacenter AI and server roadmap, affecting procurement and architecture decisions. Tech professionals must assess implications for inference cost, memory capacity, and future server performance planning.
Dossier last updated: 2026-06-01 20:11:26
Intel unveiled the Crescent Island GPU at Computex 2026, offering configurations with up to 480 GB of VRAM and built on its Arc Xe 3P architecture. Positioned as a high-memory accelerator, Crescent Island extends Intel’s graphics roadmap beyond Panther Lake integrated GPUs and targets AI, data-center inference, and content-creation workloads that demand very large memory pools. The design emphasizes large unified memory capacity rather than just raw shader throughput, signaling Intel’s push to compete in specialized accelerator markets dominated by GPUs from NVIDIA and AMD. This matters for cloud providers, AI developers, and workstation makers deciding on platform memory limits and ecosystem support for Intel’s GPU stack.
Intel plans to ship its Crescent Island AI GPU in limited quantities by the end of this year, targeting inference workloads with a cheaper, air-cooled design that uses LPDDR5 memory rather than high-bandwidth memory (HBM). Kevork Kechichian, head of Intel’s data center group, said the chip focuses on cost and thermal efficiency to challenge Nvidia and AMD, and follows Intel’s pullback from prior training-focused efforts. Intel aims to manufacture the new GPU in-house to reduce costs and is evaluating export-control-compliant variants for China. The move is part of CEO Lip-Bu Tan’s turnaround strategy and comes amid a broad semiconductor rally and renewed investor confidence.
Intel previewed its data-center AI inference GPU, Crescent Island, ahead of COMPUTEX 2026, confirming a late-2026 launch. The PCIe AIC card targets token-efficiency (Token/W) and supports a wide range of numeric formats from native FP4/MXFP4 up to FP64, with up to 480 GB of LPDDR5x memory and a 350 W power envelope. Crescent Island runs on Intel's Xe3P GPU architecture, which Intel says will be deployed across PCs, data centers, edge, and workstations — including a next-generation PC chip. Intel also emphasizes an open-source software stack and ecosystem support for the card, positioning it for AI inference workloads and broader platform integration.
Intel announced it will ship a new Xeon server processor family codenamed Diamond Rapids in 2027, built on an evolved Intel 18A-P process. The design keeps a large-die plus modular layout with a scalable SoC architecture and uniform memory latency. Diamond Rapids increases core counts by about 50%, targets stronger single-thread performance for high-end IaaS, doubles memory bandwidth via more channels and higher supported frequencies, and adds PCIe Gen 6 support. The improvements aim to boost datacenter compute density and I/O throughput for cloud and enterprise workloads, signaling Intel’s continued roadmap push in process, packaging and platform capabilities ahead of next‑generation server deployments.