Intel Unveils D270: A GPU Without Ray Tracing and Up to 480 GB of LPDDR5X for Inference

At the Open Compute Project 2026 summit, Intel officially introduced its D-series of graphics accelerators. The first model, the Intel GPU D270, was previously known as Crescent Island. It is built on the Xe3P architecture and carries up to 480 GB of LPDDR5X memory. Intel removed the 3D graphics and ray tracing blocks from the die. What remains is a pure compute engine for running AI model inference in server racks.

What Was Removed and What Remains

Officially, the D270 is still a GPU, but you can't game on it, and Intel doesn't hide that. Raster pipelines and RT cores exist for rendering. During model inference they sit idle while still taking up die area and power budget. Remove them, and both can go to matrix compute and memory controllers. Intel calls the card an inference accelerator, and the whole design serves that goal.

Xe3P extends the Xe family, which underpins the company's consumer Arc cards and its server products. The first materials give no detailed figures on the number of execution units, TDP, or performance in TFLOPS or TOPS. Those will have to come from Intel itself and from independent tests.

Betting on LPDDR5X Instead of HBM

The D270's main feature is its memory. Instead of HBM, which remains scarce and expensive, Intel chose LPDDR5X. It offers lower bandwidth per stack, but it is cheaper, draws less power, and allows a very large capacity. When Intel announced Crescent Island in 2025, the figure was 160 GB. Now the release headline cites a ceiling of 480 GB. The source material does not disclose how that number is reached (a single board, several modules, or a separate configuration).

For inference, capacity often matters more than peak speed. Model weights and the KV-cache for long contexts have to live somewhere. As a rough estimate, 480 GB is enough for a model of around 300–400 billion parameters in 8-bit format, with headroom for the cache. That avoids splitting the model across several accelerators with expensive chip-to-chip interconnects. This is an estimate, not a specification stated by Intel.

The trade-off is clear. In workloads limited by memory bandwidth, such as token generation at small batch sizes, HBM accelerators like the AMD Instinct MI350 with 288 GB of HBM3E will be faster. The D270 targets scenarios where total cost of ownership, power consumption, and capacity matter most.

The Market the D270 Is Aiming At

Inference is becoming the biggest line item for operators of AI services: training happens once, while queries keep coming. Competitors are already splitting their product lines by task. Nvidia, for example, showed the Rubin CPX with 128 GB of GDDR7 for the context-processing phase. Intel is heading the same way: a specialized chip with inexpensive memory instead of a universal monster.

Key specifications based on available data:

• Name: Intel GPU D270 (code name Crescent Island)
• Architecture: Xe3P
• Memory: LPDDR5X, up to 480 GB
• Purpose: AI inference in data centers
• Removed: 3D graphics and ray tracing blocks
• Presentation: OCP 2026 summit

What to Check Before Planning Purchases

The D270 has no public data yet on memory bandwidth, TDP, or price, so it is too early to compare it with competitors. First, find out the real numbers for tokens per second per watt and per dollar. Test your own models and your own context sizes, not marketing benchmarks.

The second question is software. The maturity of support for PyTorch, OpenVINO, and popular engines like vLLM will determine how long it takes to migrate a workload. The third is timing: when samples will reach customers, and in what form the card will ship in racks compatible with OCP standards. If Intel fills these gaps, the D270 could become a sensible alternative for large models with long contexts, where HBM accelerators are too expensive.