Intel Crescent Island GPU D-Series aims for 1.3 PFLOPS FP8

The GPU D-Series is the official brand for the chip Intel had been calling Crescent Island all year. The first product is the D270. It is built on Intel’s Xe3P architecture, and although Intel calls it a GPU, it is really a number cruncher: Intel stripped out the graphics and ray-tracing hardware and spent that silicon on compute instead. You can’t game on it.

The chip has 32 Xe3P cores grouped into four slices of eight. Together they carry 256 vector engines for general math and 256 XMX matrix engines for the heavy lifting in AI. Each core has 512KB of fast local cache, 16MB across the chip, and a shared 32MB L2 cache ties the slices together. It handles number formats all the way from FP4 to FP64, so it can run both very compressed AI models and precise scientific work.

Intel says the D270 can reach a peak of 1.3 PFLOPS in FP8. A PFLOP is a quadrillion calculations per second, and FP8 is a compact 8-bit number format that AI models often use to run faster. Memory bandwidth, how quickly the chip can pull data from its memory, is 1.5TB/s.

Treat both as ceilings, not benchmarks. These are theoretical peaks, and comparing them with rival cards only works when the precision, sparsity settings and software match. Independent tests have not been published yet.

For scale, the gap with Nvidia’s Rubin is huge on paper, but Rubin lives in a rack of 72 liquid-cooled chips that likely draw kilowatts each. The D270 is rated at 350W, about what a high-end gaming graphics card uses, and it is designed for air cooling.

The most interesting choice is memory. Most AI accelerators use HBM, a stacked, very fast and very expensive type of memory. Intel went with LPDDR5X, the same family of memory found in laptops and phones. It is slower, but cheaper, more power efficient and available in big amounts.

Intel’s own PCIe card ships with 160GB. Partner designs can go up to 480GB. That capacity is the point. Intel says one card can hold models of up to 300 billion parameters, and a server with eight cards can serve models of more than a trillion parameters.

Big memory also helps with long prompts. Before a model answers, it has to read and process everything you sent it, a stage called prefill, and keep a running memory of that context. Intel tuned the D-Series for exactly this work. The trade-off is bandwidth: for workloads that constantly shuffle data, HBM chips will still pull ahead.

The D270 is a standard PCIe 5.0 card that fits existing air-cooled rack servers with up to eight GPUs, much like Nvidia’s HGX systems. Intel pitches it for large language models, long-context work, several AI services running side by side on one server, and robotics and other tasks where fast responses matter.

Software support includes SGLang, llm-d and vLLM, popular tools for serving AI models, plus Nvidia Dynamo for coordinating work across mixed hardware. Intel’s shared Xe software base is meant to let developers move code between its laptop chips, graphics cards and data center parts.

Intel’s Anil Nanduri, who runs AI products in its data center group, called inference “the defining AI workload.” Customer sampling is underway with server makers, cloud providers and software firms. Intel has not shared pricing or a broad launch date.

The first partner plans are already public. Maxsun said at Intel’s 2026 LOEM Summit that it will build three D-Series products: a Turbo model for workstations, a Fanless model for servers and the CRI Mini PC for edge AI computing. It has not shown designs, specs, prices or dates, or said whether the mini PC uses the full D270.

You won’t buy a D270, but you may use one without knowing it. Cheaper inference hardware is one way AI features in apps, customer service bots and office software get faster or stay free. Chips that run in ordinary air-cooled servers also make it more practical for hospitals, banks and schools to keep sensitive data on their own machines rather than sending it to a cloud provider.

There is a bigger industry angle too. Nvidia dominates AI chips, and its prices reflect it. A credible lower-cost option from Intel, a company whose earlier Ponte Vecchio data center GPUs flopped, would give buyers more bargaining power.