Intel unveils GPU D-Series for enterprise AI inference

Intel has officially branded its upcoming enterprise AI inference accelerator, previously known as Crescent Island, as the GPU D-Series. The new product line is designed specifically for agentic AI workloads and server environments, utilizing an enhanced version of the company's Xe3 architecture dubbed Xe3P. The flagship model, the Intel GPU D270, is built as a 350-watt, air-cooled PCIe 5.0 x16 card, allowing for deployment in standard server infrastructure without the need for specialized liquid cooling systems.

The hardware configuration of the D270 includes 32 Xe3P cores and 256 XMX matrix engines. Intel has redesigned these engines to feature a 16-deep processing pipeline, a significant departure from the four-deep systolic design found in standard Xe3 architecture. Each core is equipped with 1MB of register file space and 512KB of L1 cache, supported by a 32MB unified L2 cache. The silicon is dedicated entirely to inference tasks, omitting traditional graphics processing elements.

Memory capacity is a primary differentiator for the D-Series. While Intel’s own reference card will ship with 160GB of LPDDR5X memory, the company is allowing original design manufacturers to equip the cards with up to 480GB of LPDDR5X. This configuration utilizes a 1.5TB/s memory bandwidth, which Intel aims to leverage for handling long-context and agentic tasks. The architecture supports a wide range of data formats, spanning from FP4 to FP64, with a peak performance rating of 1.3 PetaFLOPS for FP8 matrix operations.

Intel positions the D-Series as a power-efficient alternative to higher-wattage accelerators currently on the market. By targeting a 350-watt thermal design power, the company intends to provide a more manageable option for data centers compared to competitors that require significantly higher power envelopes. Intel claims that a single card can support models with up to 300 billion parameters, while an 8-GPU cluster is capable of running trillion-parameter models.

Software support remains a critical component of the rollout, with Intel leveraging its existing Xe software foundation to ensure compatibility with established AI frameworks. Linux driver development is currently underway, with the company expecting robust upstream support by the time the accelerators reach scale. While customer sampling has begun, Intel has not yet disclosed pricing, general availability dates, or independent performance benchmarks for the new series.