Intel's GPU D-Series: 480GB of Air-Cooled Inference That Fits Any Data Center Rack

by Geth Author

Intel gave its previously codename-only AI inference accelerator a real product name at the Open Compute Project Global Summit in October 2026: the GPU D-Series, built on its new Xe3P architecture.

The card is positioned squarely at the inference market — the side of AI that runs models in production rather than training them — and its defining characteristic is simplicity of deployment: it fits in a standard PCIe slot and cools with air. No liquid cooling. No proprietary rack. No infrastructure overhaul.

What it ships with

The GPU D-Series spec breaks down as follows:

  • Up to 480GB of LPDDR5X memory — nearly tripled from the 160GB Intel described at OCP 2025
  • 1.5 TB/s of memory bandwidth
  • 1.3 petaflops of FP8 compute
  • 350-watt power envelope
  • Standard PCIe form factor, air-cooled

A single card handles models up to 300 billion parameters. An eight-card system scales to models above one trillion parameters.

Why skipping HBM matters

One of the more interesting architectural choices here is what Intel left out: HBM. The D-Series uses LPDDR5X instead of the high-bandwidth memory found in Nvidia's H100 and H200 cards.

For inference workloads — where you're loading model weights and running forward passes, not shuffling massive gradient tensors — the bandwidth difference matters less than it does in training. LPDDR5X is cheaper, more widely available, and doesn't require the specialized handling that HBM does. The result is a card that data centers can actually order in quantity and slot into existing racks.

Intel is betting that most of the inference market doesn't need H100-style infrastructure. If that bet is right, the D-Series could undercut the cost of inference at scale significantly.

Current status

The card is still in early customer sampling. Pricing and a general availability date haven't been disclosed. Intel described it as in "bring-up" — working but not yet ready to order at scale.

The OCP reveal was nonetheless deliberate: Intel is showing hyperscalers and large enterprises that the hardware exists and is ready to evaluate. The company also announced Xeon-based rack designs at the same summit, many built on the Open Rack v3 standard, framing the D-Series as part of a broader open infrastructure push.

The D-Series is Intel's clearest shot yet at the growing inference market, where Nvidia, AMD, and Google's TPUs currently dominate. Whether the specifications hold up under real-world workloads — and at what price — will determine whether it lands.