At the Computex 2026 exhibition in Taipei, Intel announced the latest plans for its Data Center Group (DGC) and officially released Crescent Island, a new generation GPU platform for artificial intelligence data centers, focusing on high memory capacity and energy efficiency for inference scenarios. Intel said that with the rapid expansion of artificial intelligence applications, data center needs are shifting from traditional loads to AI-centric, and the company is responding to the new generation of workloads by improving performance per watt, single-core performance, rack-level core density, and memory bandwidth.
Intel believes that current data centers are gradually evolving from traditional architectures dominated by x86 CPUs to AI frontline infrastructure that requires GPU accelerated training and inference. The company expects that over the next five years, data center loads will be roughly 50:50 split between traditional and AI tasks, with the majority related to AI inference. To this end, Intel is advancing simultaneously on the CPU and GPU sides: the CPU has been introduced in previous separate conferences, but this time it will focus on the Crescent Island GPU designed for AI acceleration, especially inference scenarios.
Crescent Island is based on Intel's Arc Xe 3P architecture, which is also used in the current Panther Lake integrated graphics. It is Intel's latest generation of high-performance graphics and computing platforms for data centers. Intel says this is currently one of its most powerful data center GPUs, offering up to 480GB of video memory capacity on a single card, and is clearly targeted at AI and high-performance computing (HPC) workloads that require large-scale models and data sets.
Unlike many high-end professional GPUs that use HBM high-bandwidth memory, Crescent Island uses an LPDDR5X memory solution. Intel pointed out that this design aims to strike a balance between power consumption, bandwidth and capacity, and further improve overall energy efficiency performance. It is worth noting that compared to the 160GB graphics memory configuration announced by Intel when it first introduced Crescent Island last year, the graphics memory capacity of the officially released version has been significantly increased to 480GB, showing that its positioning for ultra-large models and data-intensive scenarios has been further strengthened.
In terms of heat dissipation design, Crescent Island adopts an air-cooling solution and is officially rated to handle a thermal design power consumption of 350 watts TDP. Intel said the card can cover the needs of a new generation of AI workloads and supports multiple data types and microscaling formats from native FP4, MXFP4 to FP64 to adapt to training and inference scenarios with different accuracy and performance requirements. The company emphasizes that Crescent Island aims to be tailored for AI inference while achieving higher performance density in overall data center power consumption and operating costs.
For users who want to focus on more affordable workstation GPUs, Intel also mentioned that its B-series Arc Pro graphics cards based on the earlier generation Xe2 architecture are still on sale, such as B70, B65 and 32GB video memory configuration models, mainly for professional graphics and entry-level AI scenarios. Compared with Crescent Island for high-end data center loads, these workstation-level products have lower prices and deployment thresholds, but there are trade-offs in graphics memory capacity and computing power.