IA en un minutoNewsBusiness and products
Intel unveils the D270, an AI inference GPU with up to 480 GB of memory; for now, only samples for customers
IA en un minuto newsroom · Editor: Jon Elgezabal
In 30 seconds
Intel now has a commercial name for its inference card: the D270 opens the D-Series family, previously called Crescent Island. The bet is on capacity, with a ceiling of 480 GB of LPDDR5X for third-party versions, and on a 350 W draw that lets it be cooled with air inside ordinary servers. For now it is not on sale: samples are with customers, and both the price and the release date are missing.
Intel is betting on memory to carve out a place in AI inference. On Monday, October 12, at the OCP Global Summit 2026, the company gave the technical details of its new Intel GPU D-Series family, known until now by the code name Crescent Island. The D270 is the first card in the series and, for the moment, the only one, according to ServeTheHome.
It is built for inference, that is, for running models that are already trained rather than training them at large scale, as the product page explains. According to Intel, it delivers 1.3 petaflops of FP8 compute and a memory bandwidth of 1.5 TB/s while drawing 350 W, and it relies on Xe3P, the company's most recent graphics architecture. Versions made by partners will be able to carry up to 480 GB of LPDDR5X; the Intel-branded card comes with 160 GB of that memory. Intel says one card can handle models of up to 300 billion parameters, and a system with eight of them, models above one trillion.
It uses the standard PCIe format and is cooled by air, so that, according to the company, it can go into data centers already in operation with no special racks or cooling systems. "Inference is becoming the defining AI workload," said Anil Nanduri, vice president for AI products in Intel's Data Center Group. WCCFtech notes that picking LPDDR5X over HBM, the high-bandwidth memory found in NVIDIA and AMD accelerators, cuts power draw, cost and exposure to a tight supply, in exchange for a lower peak bandwidth.
For now, Intel is only handing samples to customers. It says it is working with server makers, cloud providers, software firms and other companies to get validated platforms and ready-to-deploy solutions in place. According to ServeTheHome, it has not announced a launch date or confirmed that the card will arrive before the year is out, and the price is not known either.
The same outlet recalls that this is Intel's first commercial AI-focused GPU for data centers since Ponte Vecchio in 2021, after two of its successors were scrapped, Rialto Bridge in 2023 and Falcon Shores in 2025, and that Crescent Island was announced at last year's OCP Global Summit. It also sets the card against AMD's MI350P, another PCIe card, which reaches a theoretical peak of 2.3 petaflops in FP8 and draws 600 W in its most powerful configuration.
Why it matters · analysis and opinion
The fast memory fitted to the most powerful accelerators is in short supply, and Intel has chosen to go around the problem rather than fight for it. By using a cheaper memory that draws less power, it gives up part of the speed and gains capacity: more of the model and more context fit on each card, which is exactly what agents and assistants working with long documents ask for. The other strength is practical. A card that is cooled with air and fits into an ordinary server appeals to companies that want to run models on their own premises without rebuilding the machine room. The decisive part is missing: the price, the date and independent tests with real workloads. Until those arrive, it is a reasonable promise for anyone planning purchases, not an alternative that can be ordered today.
Official source: Intel · Written with the help of AI: how we make the news
