Intel confirms its next-gen 'Crescent Island' data center GPU will support up to 480GB LPDDR5X memory

Intel used its Computex 2026 stage time to pull back the curtain further on Crescent Island, its upcoming data center GPU built on the Xe3P architecture. The chip is aimed squarely at AI inference workloads, and the headline detail from the presentation is that board partners will be able to build configurations with up to 480GB of LPDDR5X memory onboard.

The 160GB figure that Intel had already confirmed represents the reference configuration, while the 480GB ceiling opens the door for AIB partners to build much denser variants, similar to how add-in board partners approach gaming GPU memory configurations.

3

What makes Crescent Island stand out from the competition is its choice of memory. While conventional AI accelerators from NVIDIA and AMD rely on expensive high-bandwidth HBM memory, Intel's chip uses consumer-oriented LPDDR5X DRAM and is designed to run in air-cooled servers rather than liquid-cooled setups.

LPDDR5X trades raw bandwidth for significantly lower cost, lower power draw, and far better supply availability, since HBM supply from companies like SK Hynix and Samsung remains a key bottleneck across the AI hardware industry. The global DRAM shortage is still ongoing, as we know, and that largely affects the data center segment as well.

Since the use of LPDDR5X significantly reduces power consumption, we get a 350W TDP for the air-cooled PCIe version. That puts it well within standard server rack power budgets, which is a real deployment advantage for data center operators who don't want to rearchitect their cooling infrastructure.

On the bandwidth side, recent leaks suggest Crescent Island may use a wide-but-slower 640-bit bus connecting around 20 LPDDR5X devices, which would yield roughly 684 GB/s of memory bandwidth using 10.7 Gbps parts. Intel hasn't confirmed this, so treat those numbers as estimates for now. Notably, Intel has not released any raw throughput figures, making direct performance comparisons with NVIDIA or AMD accelerators impossible at this stage.

3

As far as features go, Crescent Island supports a wide range of data types and microscaling formats, from native FP4/MXFP4 all the way to FP64, which gives it flexibility across inference and scientific computing workloads. Intel is also touting its oneAPI software stack for the platform, though oneAPI is significantly less widely adopted than CUDA or ROCm, and that software gap remains a real challenge for enterprise adoption.

Intel has described Crescent Island as "coming soon" and has pointed to a second-half 2026 launch. Pricing has not been announced, though the LPDDR5X approach strongly signals a cost-optimized positioning relative to HBM-based rivals. How competitive it actually ends up being will depend heavily on the throughput numbers Intel has yet to share.