Intel launches 480GB GPU D-series for AI inference
Intel's GPU D-series packs up to 480GB of LPDDR5X memory for large AI inference workloads, but pricing and general availability remain unannounced.
Intel has unveiled its GPU D-series, a data-centre accelerator with up to 480GB of LPDDR5X memory aimed at large-scale AI inference. The company says the hardware can run models with up to 300 billion parameters on one card, while an eight-GPU system can handle models exceeding one trillion parameters. 1
Why the memory matters
The D-series, formerly known by the codename Crescent Island, is designed for workloads where model weights and long prompts need to remain accessible during inference. Intel pairs the unusually large memory pool with 1.5TB/s of bandwidth and up to 1.3 petaflops of FP8 compute. 1
That is a different proposition from a conventional gaming graphics card. Intel is targeting enterprise AI services, robotics, long-context workloads and simultaneous inference requests rather than frame rates or consumer PC upgrades. The company says the accelerator uses a standard PCIe form factor, air cooling and a 350W power envelope, which should make it easier to fit into existing data-centre environments. 1
TechRadar reports that the D-series now has more memory than Nvidia's 288GB Blackwell Ultra B300 configuration, although memory capacity alone does not determine real-world inference speed. Bandwidth, software support, compute resources and the workload itself will still decide how quickly a system serves requests. 2
What Intel has actually announced
The announcement came at the Open Compute Project Global Summit, where Intel also discussed Xeon-based rack designs and its wider open-software strategy. The D-series is built on Intel's Xe3P architecture and is intended to work alongside Xeon processors, with CPUs handling orchestration, data movement and tasks that do not require GPU acceleration. 1
Intel says customer sampling is under way, but it has not announced a public retail price or a specific general-availability date. That matters: this is a significant infrastructure announcement, not a product that businesses can simply order today.
Our opinion
Intel's 480GB headline is eye-catching, but the more important detail is the trade-off behind it. The D-series is chasing capacity and power efficiency for inference rather than trying to be a faster gaming GPU. If Intel can deliver the promised software portability and keep deployment straightforward, putting more model memory on a single PCIe card could reduce the awkward compromises involved in splitting large models across hardware. The proof will come from independent performance figures, pricing and availability — none of which Intel has supplied yet.