Intel has officially presented the Intel GPU D-Series, the commercial name of the AI accelerator previously known by the codename Crescent Island. The product is optimized for inference workloads in data centers, especially large language models and agentic systems.
What happened?
At the Open Compute Project (OCP) Global Summit 2026, the company confirmed the specifications of the Intel GPU D-Series. The main model, Intel GPU D270, delivers up to 1.3 petaflops of performance in FP8 precision and 1.5 TB/s of memory bandwidth, using the Xe3P architecture.
The Intel GPU D-Series stands out for its memory capacity: ODM partner versions can reach 480 GB of LPDDR5X, while Intel's reference card comes with 160 GB. The form factor is PCIe 5.0 air-cooled with a 350 W TDP, which facilitates installation in existing servers and racks without the need for liquid cooling infrastructure.
Why it matters for AI
Unlike training accelerators that depend on high-cost and supply-limited HBM memory, the Intel GPU D-Series prioritizes memory volume and energy efficiency for inference. According to Intel, a single card can host models of up to 300 billion parameters, and an eight-GPU system can scale to more than one trillion parameters.
The silicon features 32 Xe3P cores and 256 XMX engines. Die area was reallocated from graphics and ray tracing pipelines to matrix compute, which is suitable for prefill workloads and long contexts. Samples are already available to OEMs, ODMs, and cloud providers, although price and commercial launch date have not yet been disclosed.
The Intel GPU D-Series represents the company's attempt to return strongly to the data center accelerator market for the agentic AI era, offering a more accessible alternative in terms of existing infrastructure.
Sources: official Intel GPU D-Series page, Tom's Hardware coverage, and IT之家.
By GeekikiBot
