Intel Unveils Crescent Island as the Intel GPU D-Series with 1.3 PFLOPS FP8 Performance at OCP 2026

At the Open Compute Project (OCP) Global Summit today, semiconductor giant Intel officially lifted the veil on its much-discussed artificial intelligence accelerator previously known by the codename Crescent Island, giving it a proper commercial moniker: the Intel GPU D-Series. Alongside the official naming announcement, the company provided critical performance and architectural disclosures, offering the industry a clearer picture of what to expect from its upcoming high-efficiency inference accelerators.

For the first time since the architecture began making rounds in industry discussions, Intel shared a targeted theoretical peak performance figure for the chip, boasting an impressive 1.3 PFLOPS of FP8 (8-bit floating point) compute capability. Additionally, the company disclosed the final memory bandwidth specifications for the accelerators, confirming that the hardware will deliver 1.5 TB/s of throughput. This fresh wave of technical data builds upon a steady, deliberate rollout of information that Intel has maintained regarding Crescent Island throughout the course of 2026, slowly feeding the enterprise hardware market with details about its evolving roadmap.

Industry watchers and enterprise customers have been piecing together the puzzle of Crescent Island for months, but today’s disclosures at OCP 2026 mark a significant milestone in turning the silicon codename into an actual market-ready product line. The GPU D-Series is designed to address the rapidly expanding demands of modern AI inference workloads, where energy efficiency, memory capacity, and raw throughput are paramount. By settling on a formal product family name, Intel is signaling that the accelerator is moving closer to broad commercial availability, giving data center operators and system integrators a concrete platform to evaluate for their future infrastructure deployments.

The foundational architecture of the chip has also been reinforced by today’s announcements. It is already well established that Crescent Island represents the premier implementation built upon Intel’s Xe3P graphics architecture. However, industry analysts and technical experts note that describing the chip strictly as a traditional graphics processor is a misnomer. In reality, the architecture functions as a massively parallel computing engine designed specifically for deep learning and high-performance computing workloads. The silicon intentionally omits most, if not all, of the graphics-specific functional units and rendering pipelines that would typically characterize a consumer-facing or workstation-grade GPU. By stripping away these legacy graphics components, Intel has freed up valuable die space and thermal budget to maximize raw compute density and memory efficiency for tensor operations and matrix multiplication tasks.

Complementing these computational capabilities, Intel has outlined the physical and thermal parameters that will govern how the GPU D-Series integrates into existing data center environments. The company confirmed that it will release the GPU D-Series products in a standard air-cooled PCIe form factor. This design decision is particularly crucial for enterprise customers looking to upgrade existing server infrastructure without overhauling their current cooling systems or investing in complex, expensive liquid-cooling retrofits.

The air-cooled PCIe cards will operate within a thermal design power (TDP) envelope of 350W, striking a balance between high-density compute performance and manageable thermal dissipation limits for standard server chassis. Furthermore, Intel highlighted the impressive memory subsystem accompanying the hardware, noting that the accelerators will support up to 480GB of LPDDR5X memory. This generous memory pool addresses one of the most persistent bottlenecks in modern AI inference—the ability to rapidly load and process massive large language models and complex neural network weights without stalling the processor cores.

By combining the 1.5 TB/s memory bandwidth with up to 480GB of LPDDR5X capacity, Intel is positioning the GPU D-Series as a compelling solution for memory-bound workloads where traditional high-bandwidth memory configurations might face supply constraints or cost premiums. As the OCP 2026 summit continues, these latest disclosures regarding the Intel GPU D-Series underscore the company’s ongoing efforts to carve out a dominant space in the high-growth artificial intelligence accelerator market, providing data center architects with a powerful, versatile, and power-efficient tool for deploying next-generation AI services at scale.