Intel Unveils 480GB AI Inference GPU Using LPDDR5X, Surpassing Nvidia B300 Capacity
Intel announced the GPU D-series at OCP Summit, featuring 480GB of LPDDR5X memory designed for large model inference, supporting up to 300 billion parameters on a single card.
ImportanceMaterialEvidenceE3 inspectableWrite-upQuick
At the recent Open Compute Project Global Summit, Intel officially unveiled its "GPU D-series" (previously codenamed Crescent Island) for data centers. The chip's most notable feature is its massive 480GB of LPDDR5X memory, a capacity that exceeds the 288GB of HBM3e offered by Nvidia's Blackwell Ultra B300, making it the highest-capacity data center GPU currently available.
By choosing LPDDR5X over traditional GDDR7 or HBM, Intel signals a design focus on capacity and power efficiency for AI inference workloads rather than raw bandwidth. According to official specifications, a single accelerator can support language models with up to 300 billion parameters, while an eight-GPU system can accommodate models exceeding one trillion parameters. This configuration targets long prompts, concurrent inference requests, and enterprise AI services.
To fit into existing data center infrastructure, the GPU D-series uses a standard PCIe form factor, operates within a 350W power envelope, and supports air cooling. Intel also plans to integrate the accelerator with its Xeon processors to allow CPUs to coordinate tasks and data movement. While the 480GB capacity addresses model weight storage, actual inference speed will still depend on memory bandwidth (rated at 1.5TB/s), compute resources, and software optimization. Intel has not yet disclosed retail pricing or a general availability date.