Intel details D270 AI accelerator for large-model inference

Intel shared new D-series details on October 12 at the Open Compute Project Global Summit.[1] Intel’s product page identifies the accelerator as the Intel GPU D270.[2] The D-series was previously code-named Crescent Island, which Intel first announced in October 2025.[1] Intel says it designed the hardware for large language models, long-context workloads and concurrent inference services.[3] Its listed target applications also include multi-tenant AI services, robotics and latency-sensitive workloads.[3]

Intel’s own PCIe card has 160GB of LPDDR5X memory, while partner-built cards can offer up to 480GB.[1] Intel specifies 1.5TB/s of memory bandwidth for the D-series.[3] It says the system can support models with up to 300 billion parameters on one card and more than one trillion parameters in an eight-GPU system.[3] Intel says greater memory capacity can help keep work on fewer accelerators.[1] StorageReview reported that D-series cards offer close to four times Gaudi 3’s memory capacity per card, but less than half its bandwidth.[5] That comparison makes capacity and bandwidth separate considerations.

Intel says the D-series is built on its Xe3P architecture, with 32 Xe3P cores and 256 XMX engines.[3][2] The company specifies 1.3 petaflops of FP8 compute within a 350-watt power envelope.[3] The D270 uses a standard PCIe 5.0 connection.[2] Intel says the standard PCIe form factor and air-cooled design are intended to simplify deployment in existing data centers, without specialized rack or cooling infrastructure.[3] Those design claims may matter to teams assessing whether a new accelerator fits their current infrastructure.

Intel says customer sampling for the D-series is underway, but its announcement did not provide pricing or general availability information.[3][1] Intel names PyTorch and vLLM, and says a common Xe software foundation supports portability across Intel GPU platforms.[1][3] The announcement does not include independent performance results, according to Jon Peddie Research.[6] The research firm also said the D270’s 1.5TB/s memory bandwidth could limit decoding speed on dense models.[6] Buyers therefore have published specifications and software details to assess, but not independent results in this evidence.