32-core Intel Xeon Ice Lake-SP faster than 64-core AMD EPYC in a number of HPC tasks

32-core Intel Xeon Ice Lake-SP faster than 64-core AMD EPYC in a number of HPC tasks

For Intel, Xeon Scalable processors remain the basis for building HPC and AI systems. The company once again recalled the corresponding AVX-512 instruction sets for HPC and DL Boost for inferance, as well as Optane PMem support, which changes the memory hierarchy and allows to work with giant datasets. All of this will be in the new Ice Lake-SP CPUs as well. About the architecture and new features of these CPUs, which we already told earlier. The new Sunny Cove cores together give an 18% increase in IPC. Combined with increased memory bandwidth (8 channels of DDR4-3200, up to 6 Tbytes of RAM per socket), as well as with « cheaper» AVX-512 instructions (some of them are able to run longer at higher frequencies), it gave a 1.2-1.6 times increase for popular HPC applications compared to the previous generation processors of Cascade Lake. In SPECcpu2017 (fp) the increase is exactly 38% for new systems. In addition, new products continue to support up to three Intel Speed Select profiles to optimize the CPU for specific loads. At SC20 the company shared some new details about the upcoming processors. The full specifications of Ice Lake-SP are not disclosed, but the preliminary test results are impressive. Intel cites a comparison of Intel Xeon 32-core models with AMD EPYC 7742 64-core. It is claimed that the novelties are 1.2-1.3 times faster in LAMMPS, NAMD (STMV) and Monte Carlo calculations. We are talking about Intel's two-socket system (2 × 32@2.2 GHz) versus AMD's two-socket system (2 × 64@2.25 GHz) with 256GB (16 × 16GB) DDR4-3200 RAM in both cases. Among new customers who will receive supercomputers and HPC clusters based on new processors, the company highlights the following:

The Korea Meteorological Administration (KMA) will receive the No. 5 supercomputer with 50 Pflos performance. It will be used, as it is easy to guess, for weather and climate research, modeling and forecasting. The system will be deployed in January 2021. Max Planck Computing and Data Facility will use the 9 Pflops Raven Cluster for research in physics, chemistry, biology and so on. The Japanese National Institute of Advanced Industrial Science and Technology (AIST) will use Ice Lake-SP in its new AI Bridging Green Cloud Infrastructure, which will deliver up to 850 Pflops of FP32 performance. The universities of Tokyo and Osaka will receive clusters of 2 and 2.8 Pflops respectively. These will be among the first Ice Lake-SP based systems in Japan. Oracle Cloud will have X9 cloud HPC installations based on new Intel CPUs.

Load comments