Intel details long-awaited Crescent Island AI GPU at Computex, boasts up to 480 GB of LPDDR5X to combat memory shortages

Details Shared at Computex 2026

At Computex 2026, Intel shared additional details about its next-generation data-center GPU, code-named Crescent Island, built on Intel's Xe3P GPU architecture. Intel describes Xe3P as 'built for agentic AI' and says it supports a wide range of data types, from FP4 for AI inference up to FP64 for scientific computing. Intel has not yet disclosed throughput specifications, so its compute performance cannot yet be estimated.

Power and Memory Design

Crescent Island will ship as a PCI Express add-in card with a 350W power target, putting it in a similar power/thermal class to Nvidia's RTX Pro 5000 Blackwell card. Rather than GDDR or HBM, it uses LPDDR5X memory: Intel's reference design carries 160GB, but the architecture is designed so partners can build cards with up to 480GB.

Leaks and prior analysis point to a wide, roughly 640-bit bus linking around 20 LPDDR5X devices to reach that capacity, which would require 24GB modules already available from suppliers such as Samsung; at 10.7 Gbps, that would yield about 684 GB/s of memory bandwidth.

Why LPDDR5X

Maximizing memory capacity while keeping bandwidth adequate is meant to keep more AI data resident near the GPU and reduce data movement, which could make Crescent Island a more efficient inference engine than GPUs built around lower-capacity GDDR memory. Using LPDDR5X instead of HBM also avoids competing for scarce advanced-packaging capacity and HBM supply, which could make the chip easier and cheaper for Intel to produce in volume; Intel has not said where the package will be manufactured.

Deployment and Software

Because Crescent Island is air-cooled with relatively modest power needs, it should fit into standard 4U or 5U GPU servers, which could suit on-premises inference deployments; eight cards at 480GB each would give a single server roughly 3.8 TB of local GPU memory for large models or many small AI agents.

For software, Intel is relying on its oneAPI stack, which the company describes as 'open, upstreamed, and Day 0 ready' for Crescent Island, though oneAPI has far less adoption than CUDA or ROCm. Intel has said Crescent Island is 'coming soon,' targeting a second-half-2026 launch.