Intel unveils its GPU D-series for data centers, featuring up to 480GB of LPDDR5X memory, an unusually large capacity to support demanding inference workloads

The cost of training and running large AI models may decrease as more memory-efficient hardware becomes available.

Intel has launched its new GPU D-series, specifically designed for data centers and AI inference workloads. A key feature of this new series is its exceptionally large memory capacity, with configurations offering up to 480GB of LPDDR5X memory. This significantly surpasses the memory capacity of Nvidia's Blackwell Ultra B300, which offers 288GB of HBM3e memory. Intel chose LPDDR5X, a memory type typically found in laptops, prioritizing capacity and power efficiency over the higher-speed GDDR7 used in conventional graphics cards. The D-series is engineered to support large language models with up to 300 billion parameters on a single card, and over a trillion parameters in an eight-GPU system. Intel aims for these accelerators to be easily integrated into existing data centers with a 350W power envelope, standard PCIe form factor, and air cooling. The company also plans to integrate them with Xeon processors for coordinated workload management. Pricing and a general availability date have not yet been announced.

Intel's strategic decision to utilize LPDDR5X memory for its data center GPU D-series signals a market pivot toward prioritizing sheer memory capacity for AI inference, even at the expense of peak bandwidth typically associated with GDDR7. This approach addresses the growing need to keep massive model weights accessible during inference, aiming to make high-capacity deployments more practical and economical. The emphasis on affordability and integration suggests Intel is targeting enterprises seeking cost-effective solutions for increasingly complex AI tasks, positioning itself as a challenger to established high-end inference hardware.

This move could significantly impact the competitive landscape by lowering the barrier to entry for running advanced AI models, potentially democratizing access to powerful inference capabilities. Companies that have found the memory requirements of large models prohibitive may now find Intel's D-series a more accessible alternative. Future developments to watch include how Intel's software ecosystem matures to support these high-capacity accelerators and whether this memory-centric approach gains broader industry adoption beyond specific inference workloads.