Intel has unveiled its new GPU D-series for data centers, a family of graphics accelerators built specifically for AI inference workloads. The headline feature is an unusually large memory capacity: up to 480GB of LPDDR5X, the same class of memory used in laptops, according to TechRadar.
The launch marks a direct challenge to Nvidia's dominance in AI hardware, where the Blackwell platform has set the pace. Intel says an eight-card D-series system can handle models with up to 1 trillion parameters, a figure that puts it in the same conversation as rival accelerators costing far more.
Intel GPU D-Series specs and memory
The D-series stands out not for raw compute but for how much model data it can hold close to the chip. LPDDR5X is cheaper and more power-efficient than the HBM memory used in high-end AI GPUs, which lets Intel pack more capacity per card.
That trade-off matters for inference, where the goal is serving a trained model to many users quickly rather than training a new one from scratch. Large language models are memory-hungry, and running them efficiently depends heavily on how much of the model fits on the accelerator.
- Up to 480GB of LPDDR5X memory per GPU
- LPDDR5X is the same memory type used in laptops
- An 8-card system can handle 1 trillion parameter models
- Positioned for AI inference, not training
