Intel GPU D270: Intel's inference GPU trades HBM for up to 480 GB of LPDDR5X

On October 12, 2026, at the Open Compute Project (OCP) Global Summit in San Jose, Intel gave its data-center inference accelerator a name and a spec sheet. The chip known as Crescent Island is now the Intel GPU D-Series, and its first card is the GPU D270. It's a 350-watt, air-cooled PCIe card that skips expensive high-bandwidth memory (HBM) in favor of LPDDR5X, the type of memory used in laptops and phones, in large amounts. It's a bet on a different kind of AI datacenter hardware: cheaper cards that slot into the servers companies already own.

This story is based on Intel's October 12 press release and D-Series product page, checked October 14, 2026, plus coverage from ServeTheHome, Notebookcheck, Phoronix and ComputerBase.

The short version

• What: an inference-only data-center GPU on Intel's Xe3P architecture, with 32 Xe cores and 256 XMX matrix engines.
• Memory: 160 GB of LPDDR5X on Intel's own card; up to 480 GB on partner (ODM) cards. 1.5 TB/s bandwidth.
• Compute and power: 1.3 petaflops of FP8; 350 W; air-cooled; standard PCIe 5.0.
• Capacity claim: a model of up to 300 billion parameters on one card, more than 1 trillion in an eight-card system, per Intel.
• When and how much: customer sampling is underway. No launch date or price.

What is the Intel GPU D270?

Intel first announced Crescent Island in October 2025, saying customer sampling would begin in the second half of 2026. The October 12 announcement confirms that schedule: Intel's release says the D-Series arrives "with customer sampling underway."

Last verified: October 14, 2026 (Intel press release and product page; the 160 GB figure is in the release's footnote)

ServeTheHome adds detail from Intel's materials: 32 MB of L2 cache, about 10.2 teraflops of FP64, ECC memory, and a double-wide card. Notebookcheck reports that Intel removed 3D-graphics and ray-tracing hardware, so this isn't a gaming part. All throughput figures are theoretical peaks.

Why use LPDDR5X instead of HBM?

When an AI model generates text, each new token means reading the model's weights and its growing attention cache from memory. That makes memory capacity and bandwidth, not raw compute, the limit for a lot of inference. HBM delivers far more bandwidth but is costly and in short supply. LPDDR5X is cheaper and lets Intel fit much more of it on a card.

The trade-off is speed. The D270's 1.5 TB/s is a fraction of what HBM-based flagships offer; Microsoft's Maia 300, announced the same day, lists 18 TB/s (see our Maia 300 story). So the D270 suits jobs where fitting a large model or many users' context on one card matters more than how fast each token arrives: Intel names multi-tenant AI services, long-context retrieval and coding assistants.

Intel also isn't pitching it against Nvidia's or AMD's flagships. ServeTheHome compares it with AMD's MI350P, a PCIe card with about 2.3 petaflops of dense FP8, against the D270's 1.3, at up to 600 W (see our AMD Helion story, whose enterprise rack uses MI350P). Intel's product page compares the D270 with Nvidia's RTX Pro 6000 workstation card, based on "public information and Intel internal analysis as of February 2026." That's a vendor projection, not a measured result.

When can you buy it?

Intel hasn't said. Its release and product page mention sampling to customers and partners but no general-availability date and no price. ServeTheHome wrote that server vendors "will have to look towards 2027 for the release of Intel's GPU D-series." Phoronix reported that upstream Linux driver work is well along. Treat 2027 as an informed expectation, not an Intel commitment.

On software, Intel says the D-Series uses the same Xe software foundation as its other GPUs and runs vLLM and PyTorch. It announced day-zero support for Meta's open-weight Muse Glimmer model on D-Series and Xeon.

What else happened at OCP?

The D270 was one of several inference announcements on the summit's first day.

• Intel Xeon racks: Intel showed partner rack designs on the OCP Open Rack v3 standard, with select Xeon 6+ configurations passing 46,000 cores per rack at 133.45 kW.
• AMD Helion: AMD introduced its CPU-heavy rack blueprint for AI agents (our coverage).
• Nvidia: in a blog post on October 12, Nvidia said it is sharing Vera Rubin rack and compute-tray designs with OCP through its MGX architecture, and described an updated MGX 1.2 specification that adds rack-scale blind-mate liquid cooling, a liquid-cooled 5,000 A busbar and support for racks up to 250 kW. It also said Meta is adopting its Spectrum-6 Ethernet switches. Our Vera Rubin coverage explains where that platform stands.

What it means

The flagship inference systems in the news this month, such as Vera Rubin NVL72 and Maia 300 racks, are liquid-cooled and built around HBM. The D270 targets something else: air-cooled enterprise servers with free PCIe slots. If its real-world speed per user is good enough, a company could run a large open-weight model on-premises without rebuilding its datacenter, the same pitch AMD makes for Helion's enterprise rack. That "if" is open: Intel has published no throughput results, no price and no ship date. For where the big-rack buildout is heading, see our gigawatt datacenter tracker.

What's confirmed, reported and uncertain

• Confirmed (Intel): the D-Series name and D270 card; Xe3P with 32 Xe cores and 256 XMX engines; 160 GB (Intel) and up to 480 GB (partner) LPDDR5X; 1.5 TB/s; 1.3 petaflops FP8; 350 W air-cooled PCIe; model-size claims; customer sampling; Muse Glimmer support.
• Confirmed (Nvidia): Vera Rubin rack and tray designs shared through MGX; MGX 1.2 features as described; Meta adopting Spectrum-6.
• Reported: cache, FP64 and other detail (ServeTheHome, Notebookcheck); 2027 availability (ServeTheHome's expectation).
• Vendor projection: the RTX Pro 6000 comparison.
• Unknown: price, launch date, which partners ship 480 GB cards first, and measured performance.