Intel's 480GB D-Series GPU Beats Nvidia on Memory, Not Speed

Intel's 480GB GPU bets on capacity, not bandwidth

The headline number is memory, but the interesting decision is which memory Intel picked.

At the Open Compute Project Global Summit, Intel showed the GPU D-series, the chip formerly known by the Crescent Island codename. It carries up to 480GB of LPDDR5X, the same low-power memory family found in thin laptops, rather than the GDDR7 used in graphics cards or the HBM3e stacked memory Nvidia puts in its AI accelerators.

That makes the D-series the highest-capacity data center GPU announced so far, ahead of Nvidia's Blackwell Ultra B300 at 288GB. It is a strange-looking record: Intel is winning on gigabytes using the cheapest, slowest class of memory in the building.

The trade-off is deliberate. LPDDR5X is cheaper per gigabyte and easier to cool than HBM, and Intel pairs it with a 350W power envelope, a standard PCIe form factor and plain air cooling. No liquid loops, no new rack design. The pitch is that a data center can drop these cards into what it already owns.

What 480GB actually buys you in inference

Capacity decides which models fit. Bandwidth decides how fast they answer. Intel is selling the first and hoping you forgive the second.

Intel says a single D-series card can hold large language models with up to 300 billion parameters, and an eight-GPU system can handle models beyond one trillion. That is the territory of open-weight frontier models, the kind enterprises increasingly want to run on their own hardware instead of renting cloud time.

The card is rated at 1.3 petaflops in FP8 and 1.5TB/s of memory bandwidth. Respectable, not spectacular. Nvidia's HBM-based parts move data several times faster, which matters when a model generates tokens and every layer has to be read from memory.

So the honest framing: the D-series is built for workloads where fitting the model matters more than the last millisecond. Long prompts, many concurrent requests, retrieval pipelines, robotics, enterprise assistants. Intel's Anil Nanduri put it plainly at the summit: inference is becoming the defining AI workload, and enterprises want the right performance at the right cost.

The catch: no price, no date, and a bandwidth question

Intel has not said what the D-series costs or when buyers can get one, and those two facts will decide whether it matters.

Intel confirmed neither a retail price nor a general availability date. That leaves the central question open. The whole argument for LPDDR5X is economics: if the card is not meaningfully cheaper per usable model than an HBM-based alternative, the capacity record is a footnote.

Then there is software. Intel says the D-series is compatible across its GPU platforms and designed to work alongside Xeon processors, with CPUs handling orchestration, data movement and the non-GPU parts of modern AI pipelines. Whether that translates into smooth deployment depends on how mature Intel's data center software stack is by launch, and that has been the company's weak spot for years.

Nvidia's advantage has never been just silicon. It is the years of tooling, libraries and habit that keep teams on CUDA. A cheaper card with more memory is a real opening, but only if the software does not eat the savings. The same logic applies across the AI hardware market, where supply is being pulled toward the highest bidder, as we saw when the RTX 5090 vanished from Nvidia's own store.

Availability and pricing

Intel has not announced pricing or a launch date for the GPU D-series, so there is nothing to compare against US, UK or EU list prices yet. The card is a data center product, which means it will most likely reach buyers through server OEMs and system integrators rather than retail shelves.

For European and UK enterprises, the timeline is the thing to watch: an OCP reveal usually precedes general availability by months, and export rules for high-performance accelerators can affect who can buy what and where. None of that has been detailed by Intel.

The D-series is Intel's clearest statement yet that it wants the inference market, not a fight over training clusters it cannot win. The 480GB figure is real and it is a record. Whether it becomes a product people buy depends on two numbers Intel has not shared: the price and the date.