Last weekend, Intel introduced the graphics processing architecture ( GPU ) of Ponte Vechhio which is a derivative of the Xe graphics architecture. Although it is an Xe family that also includes the GPU gaming architecture, Ponte Vecchio is not intended for home users, but is used in the data center for tasks related to artificial intelligence (AI) and high-performance computing (HPC).
Xe architecture in general consists of three main designs, namely “LP” for integrated graphics in the processor, “HP” for the enthusiast segment, and data center / AI, and HPC for high-performance computing.
Intel also claims the Ponte Vecchio is the company’s first Xe-based GPU that is optimized for HPC and AI workloads. According to SVP, Chief Architect, and General Manager of Architecture Intel, King Koduri, Ponte Vecchio was presented to align the GPU with other computer components that have diverse architectures.
“The workload of HPC and AI requires a diverse architecture, including CPUs, GPUs and FPGAs, to NNPs that process deeper learning,” Raja said in a written statement Thursday (11/21/2019). Behind the architecture, Ponte Vecchio is manufactured with 7 nanometer (nm) fabrication technology.
This GPU is also designed with the latest 3D chip packaging technology from Intel called “Foveros”. There are also other technologies that accompany this GPU, such as the design of multi-die interconnect bridges (EMIB), to the ability to interconnect with Computer Express Link .
For its debut, Intel boasted that Ponte Vecchio will be exhibited when the GPU is planted in the Supercomputer “Aurora” at the Argonne National Laboratory, USA in 2021. There are as many as six Ponte Vecchio GPUs that will support the supercomputer, along with two scalable Xeon processors “Sapphire Rapids” which are integrated with one of the latest programming models, oneAPI.
Intel boss Bob Swan said Xe-based GPU products for data center computers are likely to be launched in 2021. He continued, GPUs with Xe architecture for mainstream gaming computers will be around 2020, but it is likely to be designed with 10nm fabrication technology, not 7nm as for data centers .