Computex 2026: Intel launches Crescent Island GPU with up to 480GB VRAM

At Computex 2026 today Intel unveiled its plans for Data Center Group (DGC) and how it is looking to expand its operations to satisfy the modern workload demands. The company has explained how things are changing now, due to the rise of AI, compared to earlier data center requirements. As such it has expressed the need to maximize performance-per-watt, per-core performance, as well as the number of cores per rack, and memory bandwidth.

Team Blue has also acknowledged the need for accelerator (GPU)-coupled training and inferencing nowadays, alongside the predominantly x86 CPU-only approach. Essentially, foundational data centers are now turning into AI training frontiers. In just five years, the company is expecting a 50% split in AI and traditional data center workloads, with the majority of it related to AI inferencing.

We already covered Intel's CPU side progress in its dedicated article. In this one we look at Crescent Island GPUs, Team Blue's new powerful data center graphics card for AI acceleration, mainly inferencing.

Crescent Island is based on the company's Arc Xe 3P architecture which lies inside current Panther Lake iGPs as well. This is Intel's latest, most powerful card and it can pack up to 480 GB of VRAM capacity. Unlike typical high-end professional GPUs which rely on HBM for improving power efficiency, the Intel GPU here has LPDDR5X. While stock specs have the card at 160GB, ODM-branded partner cards, similar to how we have it for gaming GPUs from AIBs, can pack three times as much at 480 GB.

Cooling on the unit is done by an air cooler that can handle a TDP of 350 watts. Intel says that these cards can deal with next-generation AI workloads and come with support for a wide range of datatypes and microscaling formats, from native FP4/MXFP4 to FP64, and more.

If you are interested in more affordable workstation GPUs from Intel, check out the B-series PRO cards based on the older Xe2 architecture.