Intel’s Crescent Island GPU Receives D-Series Name and Supports Up to 480 GB of LPDDR5X Memory
Intel has recently launched its new “Crescent Island” D-Series GPUs, marking a significant step in the realm of artificial intelligence accelerators. Following a series of stalled initiatives in the past, this unveiling introduces a dedicated AI accelerator tailored for server environments, reflecting Intel’s commitment to advancing computing capabilities.
The D-Series GPU is equipped with 32 Xe3P cores, complemented by 256 XMX engines, and boasts an impressive 32 MB of unified L2 cache. With the capacity to accommodate up to 480 GB of LPDDR5X memory, it offers formidable memory bandwidth of up to 1.5 TB. Intel positions the GPU to achieve up to 1.3 PetaFLOPS of FP8 performance, supporting a wide range of data formats from FP4 to FP64. This versatile architecture is particularly beneficial for handling multimodal workloads, which may involve complex processes such as video generation and understanding, due to its integrated media engine with four encoders and four decoders.
A notable feature of the Xe3P architecture lies in the redesign of its XMX engines. Transitioning from the previous four-deep systolic design, the new Xe3P uses a more advanced 16-deep processing pipeline to enhance efficiency. The support for FP4 data format alongside FP8, as well as the implementation of large register files of 1 MB per core and 512 KB L1 cache per core, further elevate its computational capabilities. This configuration comprises 256 Vector engines and 256 newly designed XMX engines, meticulously organized in groups of eight per Xe3P core. Notably, the design intentionally omits traditional graphics processing elements, as the silicon is optimized exclusively for AI inference tasks that often demand extended contextual understanding.
Designed as a 350 W air-cooled PCIe 5.0 x16 card, the D-Series GPU seamlessly integrates into conventional server infrastructure, eliminating the need for specialized liquid cooling systems. To facilitate rapid deployment, Intel is leveraging its unified Xe software foundation and compatibility with established AI frameworks. Although customer sampling has commenced, specifics regarding pricing, general availability, and independent performance evaluations remain undisclosed. The competitive landscape for AI accelerators will largely depend on the cost-effectiveness of generated tokens and the maturity of the supporting software. Intel envisions a single card capable of powering models as large as 300 billion parameters, while an 8x GPU cluster could effortlessly handle trillion-parameter models, demonstrating the ambitious potential of this new series.
#technology #business
