Intel Unveils Details of Crescent Island AI Accelerator with 480GB LPDDR5X

Intel has disclosed new details about the Crescent Island graphics processor, designed for data centers and AI workloads. The accelerator is based on the Xe3P architecture, optimized for agent-based AI models. Unlike competitors, the chip utilizes LPDDR5X memory instead of GDDR or HBM, enabling capacities of up to 480GB per card.

The reference version of Crescent Island will feature 160GB of LPDDR5X, though manufacturers can scale this up to 480GB. Preliminary estimates suggest this will require 24GB modules, which are already available from suppliers like Samsung. Memory bandwidth will reach approximately 684GB/s using LPDDR5X at 10.7Gbps. This configuration aims to minimize data movement and improve inference efficiency.

The card adopts a PCIe form factor with a 350W thermal design power (TDP), comparable to Nvidia solutions like the RTX Pro 5000 Blackwell. With air cooling and moderate power consumption, the accelerator integrates seamlessly into standard 4U or 5U servers. A server equipped with eight such cards could offer up to 3.8TB of local GPU memory, enabling the deployment of large models or multiple small AI agents within a single system.

For managing distributed computing, Intel recommends its oneAPI platform, which the company states is ready for Crescent Island "from day one." While oneAPI’s ecosystem trails CUDA or ROCm in popularity, Intel highlights its openness and compatibility with the new product. Commercial launch of the accelerator is expected in the second half of 2026.