Intel’s 4th Gen Xeon Scalable processors were originally slated to enter mass production in the latter half of 2021. However, poor yields delayed it to the first quarter, and then the second quarter of 2022. According to industry sources, the Eagle Stream platform featuring the Sapphire Rapids-SP CPUs will enter volume production and begin shipping to clients in the third quarter. Based on a chiplet (tiles in Blue terminology) design, these server processors will pack up to 56 cores across four dies (14 x 4), making it the first Intel design to achieve the feat.
In addition to leveraging the faster DDR5 memory, Sapphire Rapids will also come with support for PCIe Gen 5 and CXL 1.1. Furthermore, there will be a few variants with on-die HBM memory for bandwidth-intensive workloads. These chips will try to solve the same problem as AMD’s Milan-X, albeit in a different way.
The HBM variant of Sapphire Rapids will feature four HBM2e dies in addition to the 14-core compute tiles. Each memory stack will be 8-hi resulting in a total of 32 HBM2e chips for the entire SOC. We can expect an overall bandwidth of around 2.5 TB/s (300GB/s+ per stack) from these on-die packages. The HBM2e memory on the 4th Gen Xeon Scalable processors can be configured in one of the three following ways: