Intel details Ponte Vecchio-Xe data center GPU

Under the name Ponte Vecchio, Intel is working on a GPU for data centers and supercomputers. It is a variant of the Xe GPU that consists of 7nm chiplets. Intel will make nodes with six GPUs and two Xeon CPUs and connect everything together with the CXL interconnect.

Intel will use the Ponte Vecchio GPU in Aurora, the American exascale supercomputer announced earlier this year. Intel has now announced that the Ponte Vecchio GPUs consist of several 7nm chiplets. Those chiplets will be connected via Intel’s emib and the manufacturer will also use its Foveros technology for stacking chips.

Intel does not provide concrete details about the construction of the Ponte Vecchio GPU. For example, it is not clear how many chiplets the GPU consists of and which components are stacked. In order to allow different GPUs to communicate with each other, Intel makes the Xe-interconnect. It is based on the CXL interconnect based on PCI-e 5.0.

This interconnect can also be used to connect the GPUs to processors. For the upcoming Aurora supercomputer, Intel will create nodes consisting of six Ponte Vecchio GPUs and two Sapphire Rapids processors. The latter are Xeon CPUs that will be released in 2021 and are made on an improved 10nm process.

Ponte Vecchio is based on the Intel Xe GPU. Under that name, Intel is working on GPUs for all kinds of applications. According to AnandTech, Intel once again indicated during a presentation that there is one Xe architecture, which is adapted depending on the application. For example, the variant for data centers will focus on artificial intelligence, double precision calculations and a high memory bandwidth.

Even before the Ponte Vecchio GPU for data centers comes out, Intel will release a Xe GPU for gamers. That should happen sometime in 2020. At the end of October, Intel announced that a Xe GPU has been mounted on a pcb for the first time and that it has been successfully turned on.