Intel Details Its Crescent Island Inference GPU, Now the D-Series, and Xeon Racks for Agentic AI

Intel (INTC) shared new technical details on Monday, Oct. 12, about its Intel GPU D-series, formerly code-named Crescent Island, an inference-focused data center GPU. It also introduced new Xeon-based rack designs and pitched open software as the base for agentic AI. The news came at the OCP Global Summit after 3:30 PM ET. Intel shares were down 3.23% at $101.32 as of 3:42 PM ET, in line with a broad chip sell-off that began earlier in the day.

The GPU: what Intel disclosed

The first card in the D-series is the D270, according to ServeTheHome. It is built on Intel's Xe3P architecture with 32 Xe cores and is designed for inference, not training. Intel says customer sampling is underway but has not given a general availability date or pricing.

The design leans on large, lower-cost LPDDR5X memory rather than the high-bandwidth memory used on top Nvidia and AMD parts. That favors fitting big models on a card over raw speed. ServeTheHome says the card is not meant to match Nvidia's Blackwell Ultra, and noted that AMD's MI350P is about 40% higher in FP8 compute at 2.3 PFLOPS and up to 600 W. ServeTheHome also expects server vendors to ship systems in 2027.

Xeon racks available today

On the CPU side, Intel introduced new Xeon-based rack designs. Igor's Lab reported six partner rack systems from ASRock Rack, ASUS, GIGABYTE, MiTAC, MSI and Supermicro, available today, three of them built to the ORv3 OCP standard. Intel said select Xeon 6+ configurations scale beyond 46,000 cores per rack at 133.45 kW.

Software and strategy

Intel framed the launch around open, heterogeneous infrastructure for agentic AI, where CPUs, GPUs and networking each handle different parts of a workload. Anil Nanduri said, "Inference is becoming the defining AI workload." Intel highlighted:

Day-0 model support. Support for Meta's Muse Glimmer at launch.

Agent tooling. The Intel AI for Enterprise Agent Toolkit, deployable on Xeon today and extending to Intel GPUs, on a shared Xe software foundation.

Open standards. Intel cited 116 OCP contributions since 2011 and its work in the Open Adapter Card workgroup.

What is still missing

Every performance and capacity figure above comes from Intel or from coverage of Intel's briefing. There are no independent benchmarks, no head-to-head results against Nvidia or AMD, no named GPU customers and no price. That leaves the stock reaction a read on the broader chip sell-off, not on the announcement alone: Intel was already down about 3.4% at midday, before the 3:32 PM ET release.