Intel details the GPU D270, an air-cooled inference card with up to 480 GB

Intel announced its GPU D-series, the line it first called Crescent Island, at the 2026 OCP Global Summit. The D-series product page names the Intel GPU D270, a PCIe data center GPU for AI inference. Intel lists up to 480 GB of LPDDR5X memory per card, though its own branded card carries 160 GB.

What Intel has disclosed about the D270

The Intel GPU D270 product page (opens in a new tab) and Intel's OCP newsroom post give these figures:

• Memory: up to 480 GB of LPDDR5X per card, with no HBM.
• Memory bandwidth: 1.5 TB/s.
• Compute: 1.3 petaflops of FP8, from 32 Xe3P cores and 256 XMX matrix engines.
• Power and form factor: 350 W, air-cooled, on a standard PCIe 5.0 card.

The bandwidth and FP8 numbers come from the Intel newsroom post on open infrastructure for agentic AI (opens in a new tab). The product page itself lists neither.

Intel says it moved die area away from graphics and ray-tracing pipelines toward the Xe3P cores and matrix compute. The product page also says the D270 covers precisions from low-precision inference up to FP64. It does not list the individual formats.

What changed since Crescent Island was announced

Intel first announced Crescent Island on October 14, 2025 (opens in a new tab). That post gave 160GB of LPDDR5X, an Xe3P design tuned for performance per watt, and air-cooled enterprise servers as the target. Customer sampling was expected in the second half of 2026.

In August, at Hot Chips 2026 (opens in a new tab), Intel showed the 32 Xe cores, 256 XMX engines, up to 480GB of memory and a 350-watt air-cooled PCIe card.

The October 2026 material adds the bandwidth and FP8 figures and the D270 product name. Intel now says customer sampling is underway. A footnote in the newsroom post explains the memory numbers: the Intel-branded PCIe card has 160GB, and the design lets partners build ODM-branded cards with up to 480GB. So the 480 GB headline figure applies to partner cards, and Intel's own board stays at 160 GB.

Who the D270 is for

Intel aims the D270 at teams that serve models in production rather than train them. The product page lists agentic AI, long-context RAG, enterprise and coding assistants, document understanding, multimodal AI, diffusion models and high-concurrency serving. Its named audience includes platform architects, SaaS and GPU-as-a-service providers, and government or sovereign-cloud teams.

Memory capacity is the main pitch. Intel says one card supports models of up to 300 billion parameters, and an eight-GPU system supports more than one trillion. Intel does not say which precision those figures assume, or whether they need the 480 GB partner cards.

Air cooling matters for buyers with existing racks. Intel says the D270 runs on passive cooling in "the racks you already run", without a liquid-cooling retrofit between evaluation and deployment.

Software support for the D270

The product page says the D270 serves open models on vLLM and the open PyTorch ecosystem, with zero-day support for leading open models. The newsroom post names one model with Day 0 support across the D-series and Xeon: Meta's Muse Glimmer. It also says the D-series is built on a common Xe software foundation shared across Intel GPU platforms.

Neither page gives a driver version, a minimum vLLM or PyTorch release, or install steps. Intel has not said whether an existing vLLM deployment runs on the D270 unchanged. For reference, a standard setup is covered in this guide to serving an open model with vLLM behind an OpenAI-compatible API.

How to get access to a D270

There is no public way to buy or rent a D270 yet. The product page asks interested customers to contact Intel and offers a sign-up form for D-series news. Intel names no cloud provider, server model or partner card on the pages reviewed here.

Teams planning long-context RAG on the card can still prepare the retrieval side on current hardware, for example with a RAG pipeline on Postgres and pgvector.

Limits and open questions

• Intel gives no price, no general availability date and no named cloud or server partner.
• Intel's performance projections compare the D270 with the NVIDIA RTX Pro 6000, based on public information and Intel internal analysis as of February 2026. No benchmark numbers are published.
• The product page's memory capacity comparison is made at comparable bandwidth, without naming the other product.
• Intel's pages do not say whether "D270" names the 160 GB Intel card, the partner cards or both.
• The parameter counts per card come without a stated precision or memory configuration.
• The 1.5 TB/s bandwidth figure appears in the newsroom post only.