Intel GPU D270 is a data-center accelerator designed for AI inference—the work of using a trained model to produce responses or predictions. Intel positions it for long-context and concurrent workloads, where accelerator memory and the speed of moving data matter alongside compute capacity. Intel’s product page identifies D270 as a product in the GPU D-Series; Crescent Island was its former codename.
Intel GPU D270 specifications
Intel lists 32 Xe3P cores, 256 XMX matrix engines, PCIe 5.0, 350 W power and air cooling on the product page. At the OCP Global Summit, Intel reported up to 1.3 PFLOPS of FP8 compute and 1.5 TB/s of memory bandwidth. Intel’s October 13, 2026 announcement also specified different memory configurations for its branded card and partner designs.
| Specification | Intel GPU D-Series / D270 |
| Architecture and compute | Xe3P; 32 Xe3P cores; 256 XMX matrix engines |
| Memory on Intel’s product page | LPDDR5X, up to 480 GB |
| Card configurations in Intel’s announcement | 160 GB on Intel’s branded PCIe card; partner ODM cards configurable up to 480 GB |
| Peak FP8 compute | Up to 1.3 PFLOPS, reported by Intel |
| Memory bandwidth | 1.5 TB/s, reported by Intel |
| Power and cooling | 350 W; air-cooled |
| Interface | PCIe 5.0 |
The memory figures describe different levels of the product offering: Intel’s page lists a maximum, while its announcement gives a capacity for its branded PCIe card and a maximum for partner ODM configurations. For teams choosing hardware, the card maker and configuration matter; the 480 GB maximum is not the stated capacity of Intel’s branded card.
FP8 is an eight-bit floating-point format used in AI computation. PFLOPS expresses a peak rate of floating-point operations, while memory bandwidth measures how quickly data can move between memory and the processor. Intel’s figures describe its stated specifications, not a measured workload result.
The workloads Intel is targeting
Intel describes D270 as an inference accelerator for long-context AI, concurrent sessions and agentic AI. Its listed use cases include retrieval-augmented generation (RAG), which brings relevant documents into a model’s response process, along with coding assistants, document understanding and multimodal AI.
Intel also says a single card can support models with up to 300 billion parameters, while an eight-GPU system can support more than one trillion. Those are Intel’s capability claims. The company’s product page names vLLM and the open PyTorch ecosystem; its FAQ also lists SGLang and OpenAI Triton.
Intel says it shifted die area away from graphics and ray-tracing pipelines toward Xe3P cores and matrix compute. That design and its listed workloads put D270 squarely in the data-center inference category, rather than positioning it as a consumer gaming GPU.
At the OCP Global Summit on October 13, 2026, Intel said customer sampling was underway.
