Intel Announces D270 GPU with 480 GB LPDDR5X for Agentic AI Inference

Intel has introduced the Intel® GPU D270, an inference-optimized data center GPU specifically engineered for the demands of agentic AI workloads. The card is built to handle long contexts, multistep workflows, and high concurrency by prioritizing memory capacity and prefill efficiency over traditional graphics capabilities.

Key Specifications

• Memory: Up to 480 GB LPDDR5X per card, enabling whole-context reasoning without relying on High Bandwidth Memory (HBM).
• Compute: 32 Xᵉ3P cores and 256 XMX matrix engines, with die area reallocated from graphics and ray-tracing pipelines to accelerate inference.
• Power: 350W air-cooled design, allowing deployment in existing racks without liquid-cooling retrofits.
• Connectivity: Standard PCIe 5.0 form factor, compatible with existing server infrastructure.

Target Workloads and Software

The D270 is designed for production AI inference scenarios such as agentic fleets, long-context retrieval-augmented generation (RAG), enterprise AI assistants, and coding assistants. It supports native low-precision inference up to FP64 scientific computing. Intel highlights zero-day support for leading open models served on frameworks including vLLM, SGLang, OpenAI Triton, and the open PyTorch ecosystem.

Infrastructure Implications

Intel positions the D270 as a solution for scaling inference within current power and rack constraints. By using standard PCIe and air cooling, it aims to maximize tokens per rack and watt, offering an alternative to proprietary networking or bespoke liquid-cooled systems for enterprise, SaaS, and sovereign-cloud deployments.