Nvidia’s time as the only AI hardware game in town may be coming to an end. Intel has announced its next-generation Crescent Island inference-optimized GPU based on its Xe3P architecture. This GPU will have up to 160GB of VRAM and is designed to accelerate AI inferencing workloads. The only downside is that it’s not shipping until the second half of 2026.
Until recently, if you wanted a powerful AI supercluster, you needed to pack it full of Nvidia GPUs—ideally, ones from the latest Blackwell generation. However, while Nvidia’s cards may still rule the roost in training, there are a lot of companies looking to produce leaner, cheaper, and more efficient inferencing hardware using GPUs and ASICs, and Intel’s new design could well compete in that space towards the end of next year.
Credit: Intel
Crescent Island utilizes the next-generation Xe3P Celestial micro-architecture, which is heavily optimized around performance-per-watt to maximize efficiency and extract the most value out of the AI workloads these GPUs could be employed to crunch. With 160GB of LPDDR5X apiece, these GPUs should be fast and efficient, with plenty of scope for supporting large language models.
That performance-per-watt also extends to cooling, with Intel’s GPUs reportedly being merely air-cooled, cutting down on power and water usage, per Phoronix. Intel is working to optimize the open-source software stack that the GPU will run when it debuts toward the end of next year.
But that will put it up against next-generation Nvidia and AMD GPUs. Nvidia is set to launch its Vera Rubin architecture in the second half of 2026, offering potentially big upswings in performance and efficiency. AMD has also announced its MI450 range of GPUs, which will be equally (if not more) capable than Vera Rubin if the releases are to be believed.
Either will offer stiff competition for whatever Intel puts out. They’re all intriguing GPUs, but we’ll need to hear more before we can estimate how they might stack up.