Intel GPU D Series AI Inference: Crescent Island Gets Real

Intel GPU D series AI inference is the new search everyone in data center hardware is typing this week. At the Open Compute Project (OCP) Global Summit, Intel gave its inference chip, code-named Crescent Island, a real product name. Customer sampling has started, while price and launch date are still secret. Here is what Intel confirmed and what is still missing.

Intel calls the family the GPU D-series. According to the Intel newsroom announcement, the card is built only for inference, the job of running trained models to answer requests.

The headline numbers are simple. Intel quotes up to 480 GB of LPDDR5X memory, 1.5 TB/s of memory bandwidth and 1.3 petaflops of FP8 compute. All of that fits in a 350-watt power envelope.

“Inference is becoming the defining AI workload,” said Anil Nanduri, vice president of AI Products and GTM in Intel’s Data Center Group. In his view, companies now need the right performance at the right cost, not just raw speed.

The first model in the family is the Intel GPU D270. According to XenoSpectrum, which read Intel’s product page, it uses 32 Xe3P cores and 256 XMX matrix engines.

It connects over PCIe 5.0 and comes as a standard, air-cooled PCIe card. As a result, it can go into existing servers without liquid cooling or a special rack.

Software support is the other half of the story. Intel says the card runs common frameworks such as PyTorch and vLLM through a shared Xe software foundation. That foundation also covers its other GPUs, such as the workstation Intel Arc Pro B70 for AI.

Crescent Island is not new. Intel first described the chip on October 14, 2025, XenoSpectrum notes. Back then, it promised 160 GB of LPDDR5X and customer sampling in the second half of 2026.

Intel has now kept that schedule. Sampling is underway, TechEnclave reported, although sampling is not the same as general availability.

In addition, Intel says it is working with server makers, neoclouds, software vendors and AI-native companies on validated systems. Target jobs include multi-tenant AI services, robotics and other latency-sensitive work.

How does it stack up? XenoSpectrum lined up the published specs against two other air- or passively-cooled PCIe inference cards.

So Intel wins on capacity and power, while AMD wins clearly on bandwidth. Against Nvidia’s card, Intel offers more memory at a similar bandwidth level.

Intel says one card can hold models of up to 300 billion parameters. Meanwhile, an eight-GPU system can handle more than one trillion.

However, fitting a model is not the same as running it fast. XenoSpectrum points out that a 300B model needs about 150 GB at 4-bit for its weights alone. On a 160 GB card, that leaves little room for long inputs or many users at once.

Bandwidth matters too. Token generation speed often depends on it, so a bigger memory pool mostly helps throughput and batching. For that reason, independent benchmarks will decide how good this card really is.

Intel has not shared a price. It also has not given a general availability date, only confirming that customer sampling has begun.

One detail is easy to miss. The 480 GB figure is a partner option, not Intel’s own product. Intel’s branded PCIe card ships with 160 GB, and ODM partners can build cards with up to 480 GB.

Because of this, the final price will likely vary by partner and memory size. Until Intel publishes numbers, any Crescent Island Intel price you see online is a guess.

The GPU was only part of the show. Intel also displayed Xeon-based rack designs, many built on OCP’s Open Rack v3 standard. Select Xeon 6+ setups go beyond 46,000 cores per rack at 133.45 kW.

Intel wants these racks judged by how many AI agents they can serve at target latency. In that split, CPUs handle retrieval and coordination while accelerators run the models.

“Success will depend not on any single component, but on open, heterogeneous infrastructure,” said Kevork Kechichian, who runs Intel’s Data Center Group. Intel also announced Day 0 support for Meta’s Muse Glimmer model.

The D-series is a data center part, so you cannot buy one for your desk. If you want to run AI at home, compare the options in our best GPU for AI guide. For a workstation card with lots of memory, read our AMD Radeon AI PRO R9700 review.

Sources

Intel Newsroom (October 13, 2026), TechEnclave (October 15, 2026), XenoSpectrum (October 13, 2026). *Photos: Intel Headquarters in 2023 by Coolcaesar, CC BY-SA 4.0, cropped and resized; Datacenter Server Racks by Carl Lender, CC BY 2.0, cropped and resized.*