Intel revealed details about its 7nm GPU

True, they are intended for supercomputers, however, some of the developments may well fall into desktop solutions.

At SuperComputing 2019, Intel revealed the first details about the Ponte Vecchio GPUs for the demanding computing segment. In particular, these GPUs will serve as the basis of the Aurora supercomputer with processing power of one exaflops. For the release of new products will be used 7-nanometer process technology, and their official release is scheduled for 2021.

In the Xe 7-nm graphic cores for the HPC segment, the chip maker plans to use a number of proprietary technologies, including the EMIB (Embedded Multi-Die Interconnect Bridge) for connecting several crystals on one substrate and Foveros, which describes the approach to 3D-packing of silicon chips. Through EMIB, Ponte Vecchio cores will communicate with HBM multilayer memory, and Foveros will allow them to be equipped with a huge cache.

The chip maker has not yet revealed the exact structure of the new GPUs, but it is known that they will have up to 1000 Executive Units (EU) in their arsenal, connected via the XEMF bus (XE Memory Fabric) with cache and memory.

Compute eXpress Link (CXL) will be used to combine Intel Ponte Vecchio accelerators with each other and with central processors. Roughly speaking, it is a PCI Express 5.0 with additional protocol elements, adapted for the areas of AI, machine learning and resource-intensive computing.