Intel Reveals Crescent Island AI GPU Details at Computex 2026

Intel Bets on Affordability with Crescent Island

Intel used the stage at Computex 2026 in Taipei to share fresh technical details about its upcoming data centre GPU, code-named Crescent Island. The chip is positioned as a cost-effective alternative to high-end offerings from Nvidia and AMD, specifically targeting AI inference workloads rather than the more demanding business of training large models.

A Different Approach to AI Hardware

Built on Intel's Xe3P architecture, Crescent Island arrives as a PCI Express add-in card with a 350-watt thermal design power that supports air cooling. That is a notable departure from the liquid-cooled designs increasingly common amongst high-end AI accelerators. The biggest twist, though, is memory. Instead of the high-bandwidth memory, or HBM, used by its rivals, the chip relies on LPDDR5X memory, with a reference design of 160 GB and partner configurations scaling up to 480 GB.

With an estimated 640-bit memory bus and 10.7 Gbps LPDDR5X, Crescent Island would deliver roughly 684 GB/s of memory bandwidth. That sits well below the nearly 5 TB/s offered by Nvidia's older H200 GPUs. Intel is betting that the trade-off in raw bandwidth will be offset by lower cost, greater memory capacity, and reduced power consumption, making the chip attractive for inference rather than training.

Targeting Performance Per Dollar

Intel's pitch centres on what it calls optimised token economics for customers running agentic AI systems. The chip supports data types ranging from FP4 for AI inference all the way to FP64 for scientific computing. Intel CTO Sachin Katti first announced the chip at the 2025 OCP Global Summit, describing it as part of a new annual GPU release cadence. The move also reflects a wider industry pressure: a global shortage of HBM has pushed companies to look for memory alternatives, and LPDDR5X offers a route around that crisis.

Reuters reported in October 2025 that the chip represents Intel's second attempt to break into the AI accelerator market after its Gaudi chips struggled to compete with industry leaders Nvidia and AMD.

Timeline and Competition

Intel confirmed that customer sampling of Crescent Island remains on track for the second half of 2026, though the company has not yet released specific performance benchmarks. When it arrives, the chip will compete against products like AMD's MI350P with 144 GB of HBM3E and Nvidia's H200 NVL in the inference-focused segment of the data centre market. Whether performance per dollar can win over customers chasing raw speed remains the open question.