Intel Unveils "Crescent Island" GPU for Data Centers and Workstations
Intel has officially introduced its next-generation graphics card, codenamed "Crescent Island," targeting data centers and high-performance workstations. Announced at Hot Chips 2026, this new GPU is built on the advanced Xe3P architecture and is engineered to deliver robust AI inference capabilities for demanding enterprise workloads.
Key Specifications and Architecture
The Crescent Island graphics card is designed as a 350-watt, air-cooled solution in a standard PCIe form factor, ensuring compatibility with a wide range of server and workstation configurations. At its core, the GPU features 32 Xe3P cores, each equipped with 256 XMX engines, 32 MB of unified L2 cache, and up to 160 GB of high-speed LPDDR5X memory in Intel’s reference design. For customers requiring even greater memory capacity, Intel’s original design manufacturers (ODMs) can configure the card with up to 480 GB of LPDDR5X, enabling unprecedented VRAM for large-scale AI models.
This substantial memory capacity positions Crescent Island as a compelling choice for organizations looking to run frontier open-source AI models locally, reducing reliance on cloud resources and enhancing data privacy.
Innovative Xe3P Pipeline and AI-Focused Design
The Xe3P architecture introduces significant advancements over previous generations. Notably, the XMX engines have evolved from a four-deep systolic design in the standard Xe3 to a 16-deep processing pipeline in Xe3P, dramatically increasing throughput for AI workloads. The GPU also supports both FP4 and FP8 data formats, catering to the latest trends in efficient AI computation.
Each Xe3P core is provisioned with a massive 1 MB register file and 512 KB of L1 cache, supporting high-speed data access and minimizing latency. In total, the GPU integrates 256 Vector engines and 256 XMX engines, organized into groups of eight per core. Importantly, Crescent Island is purpose-built for AI inference and does not include traditional graphics processing elements, focusing entirely on accelerating machine learning and agentic tasks.
Enabling Large-Scale AI Inference Locally
Intel is positioning Crescent Island for deployment in PCs, workstations, and edge data centers where rapid AI inference is critical. For example, running a trillion-parameter model such as Kimi K3 on-premises would typically require nearly 2 TB of system memory. With four Crescent Island GPUs, each configured with 480 GB of LPDDR5X, this level of performance becomes achievable within a single workstation, opening new possibilities for enterprises and research institutions.
While such high memory configurations come at a premium due to the cost of DRAM, the flexibility and scalability offered by Crescent Island are significant. Intel has also ensured day-one compatibility with all major AI frameworks, streamlining software integration and accelerating deployment for developers and organizations.
