Intel revealed that it will launch two types of Xeon based on the Sapphire Rapids architecture, one with HBM2e memory and one without it. At SC2021, Intel confirmed a third model that will feature four HBM2e memory stacks, each with 16 GB capacity, to increase the memory capacity to 64 GB.
Intel Xeon Sapphire Rapids: models with up to 64 GB HBM2e confirmed
Intel has confirmed a Xeon Sapphire Rapids model with 64 GB of HBM2e memory, which has been rumored for some time. The Sapphire Rapids CPUs, based on the Eagle Stream platform, will feature up to 56 cores and power consumption of up to 350W. Support for PCIe 5.0 and DDR5 technology is planned for the next generation.
In combination with the Xeon “Sapphire Rapids”, Intel also confirmed the cache configurations (L1, L2) of the Ponte Vecchio HPC GPUs. Both will power several supercomputers, including the Aurora supercomputer. Ponte Vecchio has 4 MB of L1 cache and 144 MB of L2 cache per segment, for a total of 408 MB of L2 cache.
To realize this GPU, Intel will work with TSMC, which will manufacture the “Xe-Link” segment on 7nm, the “Compute Tile” segment on 5nm, and the “Base Tile” segment on the Intel 7 process node.
The Aurora supercomputer is expected to have a whopping 54,000 Ponte Vecchio GPUs and 18,000 Sapphire Rapids processors. This equipment will be deployed at Argonne National Laboratory and funded by the U.S. Department of Energy (DOE).
Intel also confirms a roadmap in which they plan a successor to Sapphire Rapids in 2022 when Ponte Vecchio will hit the market. By 2023, we will have another Xeon successor and a new Ponte Vecchio.