Intel Unveils 'Intel GPU D-Series' Data Center Inference Accelerators
Intel announced its Xe3P-architecture data center inference GPUs at the 2026 OCP Global Summit, officially branding them as the 'Intel GPU D-series' with the flagship D270 model. The D270 features up to 480GB of LPDDR5X memory and 32 Xe cores designed specifically for large language model workloads. This launch marks a significant step for Intel in the competitive AI hardware market, providing data centers with high-capacity, power-efficient memory configurations tailored for LLM inference and long-context tasks. By utilizing LPDDR5X, Intel offers an alternative approach to handle memory-bound AI workloads compared to traditional high-cost high-bandwidth memory (HBM) solutions. The Intel GPU D270 operates at a 350W TDP with a standard PCIe Gen5 AIC form factor and air-cooling support, delivering 1.5TB/s of FP8 compute and memory bandwidth. A single card can support 300-billion parameter models, while an 8-card system can scale to handle 1-trillion parameter models.
## BACKGROUND
Data center inference GPUs are specialized processors designed to execute trained AI models efficiently rather than training them from scratch. Technologies like LPDDR5X offer a cost-effective memory alternative that provides high capacity and bandwidth, which is critical for loading massive language models during multi-tenant AI serving.