Intel introduces new inference GPU built on new architecture

Intel on Monday unveiled its next-generation Data Center GPU, code-named Crescent Island, designed to run inference workloads. The new products will feature 160 GB of LPDDR5X onboard memory, characterized by relatively low power consumption.

The tech company said the new inference-optimized Data Center GPU will utilize its Xe3P Celestial microarchitecture once it hits the market in the second half of next year. Intel acknowledged that its Xe3P architecture is a performance-enhanced version of the Xe3 architecture used in the Core Ultra 300-series ‘Panther Lake’ processors for laptops and compact desktops. The architecture will also be featured for clients on a next-gen Arc family, the Arc C-Series.

Crescent Island supports a number of data types

According to Intel’s report, the Crescent Island GPU will support a flurry of data types relevant for inference workloads and cloud providers. The tech firm didn’t mention the estimated performance of the product, but there are still some hints in Intel’s press release.

The 160 GB of LPDDR5X memory suggests the use of multiple LPDDR5X devices, as it’s significantly more than one typically expects from a graphics card. The Crescent Island GPU could use an LPDDR5X DRAM IC, which features two 16-bit channels, so its total interface width is 32 bits. The highest-capacity LPDDR5X chip is 32 GB (8 GB), meaning that 20 such chips are required to equip a graphics card with 160 GB of LPDDR5X memory.

Other companies, such as Nvidia and AMD, use top-grade HBM memory for their data centers, including HBM3E. They are also already discussing HBM4 for next-gen parts, such as Rubin and MI400.

Intel revealed that Crescent Island will be air-cooled and cost-optimized. The tech company is currently working to refine its open-source software stack for the new data center by utilizing current-generation Arc Pro B-Series GPUs. The initiative follows the firm’s currently ongoing enhancements for the Arc Pro B-Series, including the ongoing Project Battlematrix Linux driver improvements, the Intel Compute Runtime, and the Intel Xe Linux driver.

The tech firm unveiled its Arc Pro family in May this year at Computex 2025, which it said has larger memory configurations and expanded software support. Intel unveiled the graphics processing units for workstations and AI inference, marking 40 years of collaboration with local ecosystem partners in Taipei, Taiwan.

“For the past 40 years, the power of our partnership with the Taiwan ecosystem has fueled innovation that has changed our world for the better. This week, we are renewing our commitment to our partners as we work to build a new Intel for the future. Together, we will create great products that delight our customers.”
–Lip-Bu Tan, CEO of Intel.

On October 9, the tech firm also highlighted Xeon 6+ (Clearwater Forest), Intel’s first server processor based on the 18A process. The processor is scheduled for launch in the first half of 2026.

Intel unveils details for its Panther Lake processor

Intel also revealed new architectural details for its upcoming Intel Core Ultra Series 3 processor, code-named Panther Lake. The new processor will be the first product built on Intel’s 18A, the firm’s most advanced semiconductor processor, and the first of its kind to be manufactured in the U.S. The firm stated that it anticipates the chips to begin shipping this year as it continues to make efforts to reassert its leadership in high-performance computing and advanced manufacturing.

Intel said it’s producing both processors and future 18A-based products at Fab 52, its new state-of-the-art facility in Chandler, Arizona. The firm confirmed that the facility is now fully operational and on track to achieve high-volume production of 18A chips later this year. The tech firm has invested more than $20 million into the new fab and hopes it will produce its chips in the U.S. for the first time in nearly a decade.