Intel’s GPU D270 Is Its Fourth Swing at AI Chips, and Memory Is the Bet

At the Open Compute Project Global Summit in San Jose on October 12, Intel finally gave its long-awaited AI inference accelerator a name and a spec sheet. The chip known for more than a year only by its codename, Crescent Island, will ship as the Intel GPU D270, the first entry in a new GPU D-Series aimed squarely at the data center inference market.

The reveal matters less for the raw numbers, which landed near industry expectations, than for what it represents. This is Intel’s first commercial data center AI GPU since Ponte Vecchio in 2021, arriving after two canceled successors and during a year in which the company’s stock has climbed 183 percent on AI-driven CPU demand. It is Intel’s fourth swing at the AI accelerator market. Whether it connects depends on a strategy that is deliberately unglamorous: cheaper memory, ordinary servers, and efficiency instead of peak speed.

A real name and real numbers

The D270 is built on Intel’s Xe3P architecture, an inference-optimized variant of its third-generation Xe GPU design. The published specifications, reported by ServeTheHome from the OCP announcement, describe 32 Xe cores and 256 XMX matrix engines running at roughly 2.5 GHz, with 32 MB of unified L2 cache and ECC support. Peak matrix throughput is 1.3 petaflops in FP8 precision, and memory bandwidth is 1.5 terabytes per second over a 1280-bit LPDDR5X bus.

The card draws 350 watts, ships in a double-wide PCIe form factor with a PCIe Gen 5 x16 interface, and is air cooled. Intel removed conventional 3D graphics and ray tracing hardware from the design, according to VideoCardz, reallocating that silicon area to matrix math. Nothing here is accidental: every choice trims cost and power in service of a single workload, running trained AI models at scale.

Memory is the whole strategy

The headline number is capacity. Intel’s own branded PCIe card carries 160 GB of LPDDR5X memory, but the design lets original design manufacturer partners build cards carrying up to 480 GB, enabled by 24 GB LPDDR5X chips. Intel confirmed both configurations at the summit, and ServeTheHome notes the 480 GB version would hold more memory than any other announced PCIe form factor accelerator.

The 480 GB figure matters because on paper it beats reported numbers for unreleased rivals: The Inference cites AMD’s MI450X at 432 GB of HBM4 and Nvidia’s Rubin at 288 GB. The catch, which Intel discloses in its own footnote, is that the big number only exists if ODM partners actually build it. Intel’s own card ships with 160 GB.

Why does capacity matter more than bandwidth for this product? During the token generation phase of inference, accelerators are typically memory bound, not compute bound: the model’s weights must live in fast memory, and bigger capacity means bigger models fit on fewer cards. Intel says a single D270 can hold models up to 300 billion parameters, and that eight cards can serve models past a trillion. LPDDR5X, the same cheap memory found in laptops, buys that capacity at a fraction of the power and cost of high-bandwidth memory, and it sidesteps the HBM supply crunch that has constrained the rest of the industry. The tradeoff shows up in bandwidth: 1.5 TB per second trails HBM-equipped cards, but Intel is wagering that for sustained inference, capacity per dollar wins.

Intel is not trying to win the FLOPS leaderboard. It is betting that data centers serving millions of inference requests care more about tokens per watt and tokens per dollar than peak throughput.

The open lane against the liquid-cooled racks

The deeper differentiator may be power and cooling rather than memory. A 350-watt air-cooled card drops into servers that are already on the floor, no plumbing required. That stands in deliberate contrast to how Nvidia sells inference: the GB300 NVL72 is a fully liquid-cooled rack of 72 Blackwell Ultra GPUs and 36 Grace CPUs, a data center renovation disguised as a product.

The closest PCIe rival is AMD’s Instinct MI350P, which peaks at 2.3 petaflops of dense FP8, nearly double the D270, but draws up to 600 watts to get there. ServeTheHome puts the D270’s theoretical throughput roughly 40 percent below the AMD card, then adds the honest caveat that comparing architectures on theoretical throughput rarely predicts real-world results. Intel is also driving an OCP workgroup to standardize the sub-rack accelerator card, the form factor Nvidia’s rack-scale approach skips entirely.

There are early signs the ecosystem is listening. Intel said systems are being built by ASUS, HPE, Dell, and Supermicro, plus several neoclouds, and that customer sampling is underway with server OEMs, ODMs, and independent software vendors. For a company that has spent years announcing AI GPUs that never shipped, partner hardware in development is a meaningful credibility signal.

Two cancellations and an interim chip

Context is everything here. Ponte Vecchio, Intel’s last data center GPU, launched in 2021. Its two planned successors were canceled along the way: Rialto Bridge in 2023 and Falcon Shores in 2025. Crescent Island was unveiled at last year’s OCP summit as an interim, cost-optimized product aimed at inference, a humbler ambition after the flagship efforts collapsed.

That history explains the rebrand. Intel is dropping the Data Center GPU Max and Flex branding used for the Ponte Vecchio and Arctic Sound era and starting fresh with the GPU D-Series. The name invites comparison to Xeon D, but that line targeted networking and edge devices and was put on a retirement glide path in 2025 with the Xeon 6 SoC series. The D-Series is its own product, and Intel clearly wants the market to forget the lineage that came before it.

What Intel still owes the market

For all the detail Intel locked down at OCP, three numbers are still missing: price, launch date, and independent benchmarks. Customer sampling has begun, but Intel has not said when general availability arrives, and ServeTheHome expects server vendors to wait until 2027 for volume. Until a benchmark lands, the D270’s competitiveness on tokens per watt, the metric Intel wants judged on, is a claim rather than a fact.

The market’s reaction was muted. Intel shares fell 3.2 percent to $101.35 on announcement day, a reminder that holders sitting on a 183 percent run are underwriting the GPU leg of the turnaround and wanted more than a spec sheet. The rally’s real engine remains the CPU business: revenue hit $16.1 billion in the June 2026 quarter, up 25.4 percent year over year, the largest of the last twelve quarters, driven by CPU demand from the AI build-out.

The next readings come soon. Intel reports third-quarter results in the coming weeks, and any future D-Series announcement that carries a price or a ship date will tell us whether the fourth swing is a product launch or another press release with silicon attached. The specs are real now. The business has to follow.