Intel announces new GPU Crescent Island, customer testing to begin next year

Intel announced on Tuesday a data center GPU equipped with 160GB of memory and high energy efficiency, adding it to the company's AI accelerator portfolio, aimed at promoting Intel's new AI strategy centered on open systems and software architecture. This GPU, codenamed "Crescent Island," is designed specifically for running inference workloads on air-cooled enterprise servers, emphasizing "power consumption and cost optimization." Crescent Island utilizes Intel's Xe3P microarchitecture, which focuses on high performance under unit power consumption, equipped with 160GB LPDDR5X memory, and supports multiple data types, providing ample operating space for large language models (LLM). Intel's announcement also noted that Crescent Island will support various data types and is positioned as "very suitable" for vendors providing tokens-as-a-service and AI inference use cases. In addition to emphasizing energy efficiency, Crescent Island will adopt an air cooling design and aim for cost optimization. Intel is currently advancing its open-source software stack through the existing Arc Pro B series GPUs in preparation for Crescent Island. Intel stated that it plans to begin providing samples to customers in the second half of 2026. However, Intel did not disclose a formal launch date—whether it will be released within 2026 remains unclear, and it is more likely that a large-scale launch will wait until 2027. No product slides, prototypes, or further technical details have been released. Intel did not announce any updates regarding "Jaguar Shores." Jaguar Shores is a next-generation GPU aimed at rack-level platforms that Intel announced earlier this year. In a press briefing last month, Intel's Chief AI and Technology Officer Sachin Katti stated that Crescent Island features "enhanced memory bandwidth" and "large memory capacity," making it an "ideal choice for token cloud services and enterprise-level inference scenarios." Crescent Island was officially unveiled at the 2025 OCP Global Summit, marking Intel's formal initiation of an annual new GPU release cadence. Just a week prior, Intel had heavily promoted the upcoming "Panther Lake" and "Clearwater Forest" CPUs, and this GPU release is part of the same series of actions. Media reports indicate that over the past two years, NVIDIA and AMD have shifted to an annual new product release rhythm, and Intel's move is also aimed at catching up. Over the past 15 years, Intel has experienced multiple failures in the acceleration chip sector, going through four CEOs, and has struggled to establish a foothold in the AI infrastructure market dominated by NVIDIA. Sachin Katti was appointed by Intel CEO Pat Gelsinger in April this year to lead the company's new AI strategy. He stated that Intel is building a new vision for the AI hardware market around "open systems and software architecture," aiming to provide "reasonably scaled and cost-effective" computing power to support future autonomous AI workloads He said: "We will build scalable heterogeneous systems to provide a frictionless user experience for agentic AI workloads, while achieving optimal performance per dollar for these workloads through an open heterogeneous architecture." Katti stated that this open strategy will provide more choices for customers and partners at both the system and hardware levels, allowing multiple vendors to participate. He added: "As we continue to bring more disruptive technologies, these new technologies can be seamlessly integrated into this open heterogeneous architecture."