Intel has officially named its previously-teased 'Crescent Island' project the D-Series GPU, a server-focused AI inference accelerator built on a new Xe3P architecture with up to 480GB of LPDDR5X memory. It's designed for data centers running large language models, not for gaming or consumer workstations.
Intel has pulled back the curtain on its long-rumored 'Crescent Island' project, officially naming it the D-Series GPU. According to TechPowerUp, the card is a dedicated AI accelerator built for server deployments rather than desktop gaming, following earlier stalled attempts at a discrete data-center GPU. The chip, internally designated D270, is built on Intel's new Xe3P architecture and packs 32 Xe3P cores, 256 XMX engines, and 32 MB of unified L2 cache.
What stands out most is the memory configuration: Intel's reference card ships with 160 GB of LPDDR5X, but the company is allowing ODM partners to scale that up to 480 GB, with Wccftech reporting that some add-in-card variants could even reach 512 GB. Bandwidth tops out at roughly 1.5 TB/s. Intel rates the card for up to 1.3 PetaFLOPS of FP8 compute, with support spanning FP4 through FP64, and notes the Xe3P pipeline uses a redesigned 16-deep XMX processing path versus the 4-deep design on regular Xe3 silicon. Notably, there are no graphics output elements on this chip at all — it's built purely for AI inference workloads like long-context and agentic tasks.
The card is designed as a 350W, air-cooled PCIe 5.0 x16 accelerator, meaning it can slot into existing server infrastructure without specialized liquid cooling, though Wccftech also reports Intel has shown off a liquid-cooled dual-chip OAM module for partners wanting denser deployments. Intel envisions a single D-Series card handling AI models up to 300 billion parameters, with 8-GPU clusters scaling to trillion-parameter models, and has announced Day-0 support for Meta's Muse Spark Open model using a four-card, 640 GB configuration.
This is squarely a data-center product aimed at competing with accelerators from AMD and NVIDIA in the AI inference space — not a GPU gamers or PC builders will see on store shelves. Intel has confirmed customer sampling is underway, but pricing, general availability, and independent performance benchmarks have not yet been disclosed.
What We Know
| Spec | Detail |
|---|---|
| Architecture | Xe3P (32 Xe3P cores, 256 XMX engines) |
| Process Node | TSMC N3P |
| L2 Cache | 32 MB unified |
| Memory | 160 GB LPDDR5X (reference), up to 480 GB in partner configs; some AICs may offer 512 GB |
| Memory Bandwidth | Up to 1.5 TB/s |
| Compute Performance | Up to 1.3 PetaFLOPS FP8 |
| TDP / Cooling | 350 W, air-cooled (liquid-cooled dual-chip OAM variant also shown) |
| Form Factor | PCIe Gen5 x16 card; OAM module for partners |
Frequently Asked Questions
Is the Intel Crescent Island D-Series a gaming graphics card?
No. According to TechPowerUp, the Xe3P chip powering the D-Series has no graphics output elements at all — it's designed purely for AI inference workloads, not gaming or general graphics rendering.
How much memory does the Intel D-Series GPU have?
The reference card ships with 160 GB of LPDDR5X memory, but Intel is allowing partners to build versions with up to 480 GB, and Wccftech reports some add-in cards may reach as high as 512 GB.
What architecture does Crescent Island use?
It's built on Intel's new Xe3P architecture, featuring 32 Xe3P cores and 256 redesigned XMX engines with 16-deep processing, built on TSMC's N3P process.
When will the Intel D-Series GPU be available?
As of the latest reports, Intel has only confirmed that customer sampling is underway. Pricing, release dates, and independent benchmark results have not yet been announced.
