Intel's new GPU packs up to 480GB of memory — but skips 3D graphics

Intel has revealed the first accelerator in its GPU D-series, previously known by the codename Crescent Island: the D270, a data-center card built for AI inference rather than conventional 3D graphics. Based on Intel’s Xe3P architecture, the accelerator can be configured with up to 480GB of LPDDR5X memory.

One of the accelerator’s most unusual features is its memory configuration. Intel’s own reference card comes with 160GB of LPDDR5X, while its manufacturing partners can develop versions with up to 480GB. Rather than using high-bandwidth memory (HBM), as many competing AI accelerators do, the D270 relies on LPDDR5X. Intel is prioritizing memory capacity and deployment costs for inference workloads, where large language models need substantial amounts of memory to store model weights and manage long conversations.

The accelerator offers 1.5TB/s of memory bandwidth and a peak FP8 performance of 1.3 petaflops. It has a 350W power target and uses an air-cooled PCIe 5.0 x16 design, allowing it to fit into conventional server infrastructure without requiring specialized liquid-cooling systems.

The GPU contains 32 Xe3P cores, 256 XMX engines for accelerated matrix operations and 32MB of unified L2 cache. Intel has removed conventional 3D-graphics and ray-tracing hardware, allowing more silicon area to be dedicated to compute and AI processing. Consequently, the D270 is not a gaming graphics card, despite being based on GPU technology.

Intel says the GPU D-series is intended for large language models, long-context inference and services that handle multiple AI requests concurrently. A single card can accommodate models with up to roughly 300 billion parameters, while a server equipped with eight accelerators can support models exceeding one trillion parameters. Intel lists potential uses including agentic AI, enterprise assistants, coding assistants, document understanding, multimodal AI, diffusion models and high-concurrency serving.

Intel says customer sampling is underway with server manufacturers, cloud providers and software vendors. The company has not announced pricing or a general availability date for the GPU D-series.