The Crescent Island project moves from speculation to concrete technical specifications. At Hot Chips 2026, Intel detailed the internal architecture of its AI inference GPU, confirming a configuration based on 32 Xe3P cores and a memory system capable of reaching 480 GB of LPDDR5X. While previous reports focused on memory capacity, this presentation highlighted the pipeline details and hardware choices that define the card's enterprise positioning.

Add AlexTech.ai asPreferred Source on Google

Xe3P Architecture and 16-Stage Pipeline

The heart of the graphics processor is the Xe3P IP, which introduces a significant modification compared to the previous generation. The XMX engines have been redesigned from a four-deep systolic design (typical of standard Xe3) to a 16-deep processing configuration. This architecture is supported by massive register files, with 1 MB per core, and a 512 KB L1 cache per core. In total, the GPU integrates 256 vector engines and 256 XMX engines, organized in groups of eight per Xe3P core. The card natively supports FP4 and MXFP4 data formats, along with FP8, optimizing workflows for modern AI models.

Leaked Intel Crescent Island PCB Surfaces with Xe3P GPU, 160 GB LPDDR5X ... — https://www.techpowerup.com/349209/leaked-intel-crescent-island-pcb-surfaces-with-xe3p-gpu-160-gb-lpddr5x-and-16-pin-power

LPDDR5X Memory and Absence of Video Output

The most distinctive choice remains the abandonment of HBM (High Bandwidth Memory) in favor of LPDDR5X. The Intel-branded Crescent Island will be equipped with 160 GB of memory, while ODM partners can configure the card up to 480 GB to maximize available VRAM. A crucial detail emerging from the presentation is that the silicon does not include graphics processing elements for visual output: the card is designed exclusively as a computational accelerator and cannot be used for rendering or video output. This makes it a purely computing component, ideal for data centers and workstations where inference is the primary workload.

Energy Efficiency and Market Target

With a TDP of 350 W and air cooling in PCIe format, Crescent Island aims for a high efficiency ratio. Intel positions it for PC, workstation, and data center edge deployments, targeting users who need rapid AI inference without the high infrastructure costs associated with high-density GPUs. The company cites running trillion-parameter models as an example, suggesting that the combination of massive memory and a token-per-watt optimized architecture is designed to make complex workloads accessible in less centralized environments. Customer sampling is expected for the second half of 2026, with commercial availability anticipated in 2027.