At the 2025 OCP Global Summit, Intel announced the launch of a data center GPU codenamed "Crescent Island," designed specifically to handle growing AI inference workloads. It also simultaneously released the Gaudi 3 rack-scale reference design, further expanding its AI accelerator portfolio.
New data center GPU dedicated to the AI inference market
The "Crescent Island" data center GPU is optimized for air-cooled enterprise-class servers. It uses the Xe3P microarchitecture and is equipped with 160GB of LPDDR5X memory. It focuses on high memory capacity and energy efficiency, making it particularly suitable for "Token-as-a-Service" providers and various inference application scenarios.
“AI is moving from static training to real-time, ubiquitous inference computing, a shift driven by ‘agent AI,’” said Sachin Katti, Intel’s chief technology officer. “To effectively scale these complex workloads, heterogeneous systems are required that leverage the strengths of different chips.”
The new GPU is expected to begin sampling to customers in the second half of 2026, and the related software stack is currently being developed and tested on Arc Pro B-series GPUs.
Gaudi 3 rack-scale reference design supports large-scale model inference
The Gaudi 3 rack-scale reference design, announced simultaneously, supports up to 64 accelerators per rack, equipped with 8.2 TB of high-bandwidth memory and liquid cooling, and is optimized for large models and real-time inference. This design allows customers to flexibly scale to full rack-scale inference performance from existing PCIe infrastructure.
The author's opinion
Intel's product updates demonstrate its strategic shift from prior emphasis on training performance to embracing the rapidly growing inference market. The Crescent Island, targeted at enterprise-grade air-cooled servers, fills a niche in the mid-to-high-end inference market, complementing the training-focused Gaudi series.
As AI applications accelerate, inference workloads are becoming a major market force. Intel is combining its Xeon 6, Gaudi 3, and new data center GPUs to create a complete end-to-end solution, attempting to differentiate itself from competitors like NVIDIA and AMD. This flexible, heterogeneous system architecture is particularly attractive to cost-sensitive enterprise customers.
