Intel has unveiled a new artificial intelligence processor specifically designed for inference workloads, marking its effort to regain a foothold in the lucrative AI accelerator market.
The new graphics processing unit, codenamed Crescent Island, is optimized for energy efficiency and aims to support a broad range of AI applications. Intel's Chief Technology Officer, Sachin Katti, announced the new chip at the Open Compute Project (OCP) Summit.
"This underscores the focus I've previously emphasized—optimizing AI for the best token economics and performance per dollar," Katti remarked.
The Crescent Island GPU leverages Intel’s newly introduced Xe3P architecture, which was first revealed at a special event just a week prior. That event also introduced the upcoming Panther Lake Intel Core Ultra series 3 processors for laptops. Xe3P represents an evolution of the existing Xe3 architecture, with enhancements focused on power efficiency and cost reduction, specifically tailored for air-cooled data centers and targeting AI inference tasks. The company stated that this architecture will support a diverse array of data types, positioning it as an ideal solution for "token-as-a-service" providers and inference-based applications.
The new GPU will feature up to 160GB of memory, but intriguingly, Intel has opted for the LPDDR5X standard rather than the high-bandwidth memory (HBM) used by competitors such as Nvidia and AMD in their high-end offerings.
These rivals are currently deploying HBM3E in their latest chips and have already begun discussing a shift to HBM4 for next-generation processors like Vera Rubin and MI400. Intel's choice, however, highlights the supply chain challenges these companies face in acquiring HBM memory chips, as demand continues to rise. HBM availability has been constrained, and prices have surged in recent months. By utilizing LPDDR5X in its new GPU, Intel may gain a cost-efficiency advantage.
Intel has not yet released detailed specifications but mentioned that customer sampling for Crescent Island is currently planned for the second half of 2026, meaning more details are expected in the coming months.
What is clear is that Crescent Island will play a crucial role in Intel's renewed attempt to capitalize on the AI boom—a trend that has already generated billions in revenue for Nvidia and other chipmakers. Intel initially missed the opportunity in generative AI and has significantly lagged behind its competitors, facing a formidable challenge to secure a meaningful share in the AI accelerator market, though it cannot afford to remain sidelined.
In fact, earlier this year, Intel's new CEO, Pat Gelsinger, pledged to revitalize the company's position in the AI market. He has implemented bold strategies, including pausing earlier initiatives like the Gaudi GPU project.
At the OCP Summit, Katti revealed that Intel plans to release new GPU versions annually, aligning with the cadence of rivals like Nvidia and AMD, as well as cloud infrastructure providers such as Amazon Web Services and Google, which develop their own AI-targeted processors.
Intel faces a steep uphill battle to stand out in an AI chip market currently dominated by Nvidia. However, Katti emphasized that the company will focus on designing chips specifically for AI inference—meaning the silicon used to run AI models, rather than train them. "Rather than trying to build for every workload, our focus will increasingly center on inference," he explained.
In a surprising twist, Nvidia itself may inadvertently aid Intel's resurgence in the AI market. Last month, after investing $5 billion in its struggling rival, Nvidia now holds approximately 4% of Intel’s shares, making it one of the largest shareholders alongside the U.S. government. Additionally, the two companies have agreed to collaborate on co-developing chips for data center servers and personal computers.
Regarding the investment, Katti noted that it will help ensure Intel's central processors will be integrated into every AI server and PC sold in the future.
