Intel Unveils 480GB GPU D-Series for AI Inference

Intel has unveiled its GPU D-series, an accelerator with up to 480GB of memory designed for AI inference. The company presented it at the recent Open Compute Project Global Summit. The product was previously known by the codename Crescent Island.

Intel says one card can support large language models with up to 300 billion parameters, while an eight-GPU system can accommodate models exceeding one trillion parameters. The company describes the accelerator as aimed at workloads including lengthy prompts, simultaneous inference requests, enterprise AI services and robotics.

The D-series uses LPDDR5X memory rather than GDDR7. Intel rates it at 1.3 petaflops for FP8 computation and specifies memory bandwidth of 1.5TB/s and a 350W power envelope. It uses a standard PCIe form factor and air cooling, and Intel says it is software-compatible across its GPU platforms.

Intel also plans to pair the accelerator with Xeon processors for task coordination and data movement. The company has not announced a public retail price or a specific general-availability date. Memory capacity, bandwidth and computing performance are separate measures; the source does not provide independent benchmarks for serving requests.