Intel names Crescent Island the GPU D-series, led by the air-cooled D270

At the Open Compute Project Global Summit, Intel’s newsroom said it was giving a product name to the inference GPU formerly called Crescent Island. The name is the Intel GPU D-series. The first card is the Intel GPU D270. ServeTheHome and Phoronix both date their reports 12 October 2026. Phoronix’s timestamp is 16:53 EDT. The newsroom post says the details were shared “today” at the summit. The text fetched from that page does not include a dateline.

The specifications Intel states

Intel’s newsroom says the D-series is built on the Xe3P architecture and combines up to 480 GB of LPDDR5X, 1.5 TB/s of memory bandwidth, and 1.3 petaflops of FP8 compute inside a 350-watt envelope. A footnote on that page says Intel’s own branded PCIe card has 160 GB of LPDDR5X, and that the design lets partners build ODM-branded cards with options up to 480 GB. The “up to 480 GB” line is the partner option. It is not the capacity Intel states for its own branded card.

The D270 product page states 32 Xe3P cores and 256 XMX matrix engines, memory up to 480 GB of LPDDR5X, a 350 W air-cooled card, and a standard PCIe 5.0 connection. It says the card is meant to keep long context resident, and it says there is no high-bandwidth memory. The product-page text opened here does not repeat the 1.5 TB/s or 1.3 petaflop figures. Those two numbers come from the newsroom post. Neither Intel page states a price or a general-availability date. The newsroom says customer sampling is underway.

ServeTheHome’s specification table also lists PCIe Gen5 x16, a GPU clock of about 2.5 GHz, 10.2 TFLOPS of FP64, 32 MB of L2 cache, and ECC. ServeTheHome presents the clock as its own derivation from the 1.3 petaflop figure and the 256 engines. Those extra lines are not in the Intel pages opened for this story. The lane count, x16, is in ServeTheHome’s table. Intel’s product page says “standard PCIe 5.0” and does not print the lane count. Phoronix repeats the core count, the 256 engines, 480 GB, 350 W, and PCIe 5.0. It does not mention the 160 GB branded card.

There is no independent benchmark in these pages. The product page says performance projections against the RTX Pro 6000 rest on publicly available information and on Intel internal analysis as of February 2026. The page text opened here does not state a projected margin, so this story does not quote one.

Why Intel says capacity is the point

Intel’s product page describes agentic inference as a workload in which agents read far more than they write: long contexts, many steps, and compute dominated by prefill. The page says the pressure shows up as memory capacity, and it specifies LPDDR5X on an air-cooled PCIe card with no high-bandwidth memory. The newsroom says the card is aimed at models of up to 300 billion parameters on one GPU, and at more than one trillion parameters in an eight-GPU system. Those ceilings are Intel’s claims. Anil Nanduri, vice president of AI products and go-to-market in Intel’s Data Center Group, is quoted: “With Intel GPU D-series, we’re focused on making high-capacity inference practical to deploy and economical to scale.”

ServeTheHome’s reading of the same announcement is that Intel is sampling cards and has not announced a launch date or a price. That matches what the Intel pages say. ServeTheHome’s further comment, that buyers may be looking to 2027, is the outlet’s conclusion. Intel’s pages do not give that year.

The racks, briefly

The same newsroom post says Intel showed new Xeon-based rack designs made with partners, many of them on the Open Rack v3 standard. It says selected Xeon 6+ configurations scale past 46,000 cores per rack at 133.45 kW of total rack power. That density figure is Intel’s. The same summit also drew rack proposals from NVIDIA and AMD, covered in a separate post, and Microsoft’s plan for the Maia 300 inference chip, covered in another.