Intel Expands AI Infrastructure for Agentic Workloads
At the Open Compute Project Global Summit 2026, Intel introduced new building blocks for data centers designed to power agentic AI applications. As AI shifts from simple queries to multi-step workflows—where agents must retrieve data, reason, and use tools—the demand for complex orchestration increases. To meet this, Intel is promoting a heterogeneous approach where CPUs, GPUs, and accelerators work in tandem through interoperable software.
Xeon Racks and Open Data Center Standards
Intel's vision centers on assigning every prompt to the most efficient resource. To support this, the company is developing Xeon-based rack solutions designed with partners, many of which follow the Open Rack v3 (ORv3) standard. This ensures that components can be easily integrated into open data center environments.
For specific configurations utilizing Xeon 6+, Intel reported a massive scale: over 46,000 cores and a total power capacity of 133.45 kW per rack. These high-density designs are intended to provide the computational foundation for agentic AI, where the CPU handles orchestration and data movement while GPUs manage specialized mathematical tasks.
Crescent Island: Targeting Memory-Intensive Inference
The GPU-D series, codenamed Crescent Island, is built on the Xe3P architecture. This series is specifically positioned for inference—running already trained models—particularly large language models (LLMs) that require long context windows. The technical specs include up to 1.3 Petaflops of FP8 performance within a 350W power envelope, paired with a memory bandwidth of 1.5 TB/s.
Memory capacity is a critical differentiator for these workloads. While Intel's own PCIe card is slated to feature 160 GB of LPDDR5X, the company is enabling Original Design Manufacturers (ODMs) to offer flexible configurations with up to 480 GB of memory. This capacity allows for handling models with up to 300 billion parameters on a single card, or over a trillion parameters in an eight-GPU system.
Software and Ecosystem Integration
To bridge the gap between model evaluation and production, Intel is focusing on a unified Xe software base. This aims to make applications portable across different Intel GPU platforms and simplify integration with existing AI frameworks. Furthermore, Intel is contributing to the Open Compute Project through the Open Adapter Card Workgroup, working to define standardized form factors for accelerators to ensure interoperability across various system designs.
Currently, the GPU-D series is in the customer sampling phase, with general availability dates yet to be announced.
Гика
