Intel Merges Habana With Xe GPUs

The company has streamlined its data-center AI roadmap, creating a new version of Falcon Shores that combines the Xe GPU architecture with some aspects of Habana’s AI design, which is delayed to 2025.

Linley Gwennap

After years of juggling too many AI-acceleration architectures, Intel is streamlining its strategy. For the data center alone, the company currently offers its Xe Max GPUs, its software-incompatible Habana Gaudi chips, and the integrated AMX unit in its newest Xeon processors. Intel’s new roadmap, however, merges aspects of Gaudi into the Xe GPU line in a product called Falcon Shores, which also brings the Habana design under the OneAPI software umbrella. But this unification won’t reach customers until 2025 at the earliest, when Falcon Shores is now scheduled for production.

The recently disclosed data-center roadmap shows an annual cadence of new mainline Xeon processors, with Emerald Rapids (which uses the same Intel 7 process as Sapphire Rapids) late this year followed by Granite Rapids (Intel 3) next year, as Figure 1 shows. The company has accelerated its first server processor based entirely on power-efficient E-cores, Sierra Forest, to enter production in 1H24 on Intel 3, followed by Clearwater Forest (on 18 A) in 2025.

After shipping Ponte Vecchio as its first Xe Max GPU, Intel had planned to follow with Rialto Bridge late this year and a subsequent GPU called Falcon Shores next year. But the new plan cancels the former and delays the latter by a year to enable the redesign with Habana technology. After Gaudi3 debuts in 2024, the company plans no further chips in that product line, with the Xe Max becoming its primary product for both supercomputing (HPC) and AI applications.