Intel has announced its Crescent Island graphics card, an inference-optimized GPU based on the Xe3P architecture. The device is designed to accelerate artificial intelligence workloads and is scheduled to begin shipping in the second half of 2026.
The Crescent Island GPU utilizes the next-generation Xe3P Celestial micro-architecture, a design heavily optimized around performance-per-watt to maximize efficiency. The card will be equipped with up to 160 GB of LPDDR5X VRAM. This large memory capacity provides substantial scope for supporting large language models and other memory-intensive AI inference tasks. The focus on efficiency aims to extract maximum value from the AI workloads the GPU will process.
The emphasis on performance-per-watt extends to the GPU’s physical design and operational requirements. According to a report from Phoronix, the Crescent Island cards will be air-cooled. This cooling method contrasts with more resource-intensive liquid cooling systems, contributing to a reduction in both power and water usage for data center operations. To support the hardware at launch, Intel is also actively developing an optimized open-source software stack that will be available when the GPU debuts.
Intel’s announcement enters a market where Nvidia’s GPUs, particularly the latest Blackwell generation, have been a primary choice for building powerful AI super-clusters for training models. While Nvidia maintains a strong position in the training sector, the market for inference hardware is becoming more diversified. A number of companies are now developing leaner and more efficient hardware, including both GPUs and ASICs, creating a competitive space for inference-specific solutions.
The late 2026 release schedule places Crescent Island in direct competition with next-generation products from other major hardware manufacturers. Nvidia is set to launch its Vera Rubin architecture in the second half of 2026, which is anticipated to deliver significant performance and efficiency improvements. Concurrently, AMD has announced its MI450 range of GPUs. Projections suggest this series will offer capabilities that are comparable to, or potentially greater than, those of the Vera Rubin architecture.
