The director of the University of Tsukuba’s Center for Computational Sciences, Taisuke Boku, shared details about the Pegasus supercomputer, HPC Wire reports. The “big memory” supercomputer is among the first to use Nvidia H100 GPUs and Intel Sapphire Rapids processors.
Built by NEC, Pegasus consists of 120 compute nodes, each equipped with one Nvidia H100 PCIe GPU and one 48-core Intel Sapphire Rapids processor (running at 2.1 GHz), providing a total of 6.5 petaflops of theoretical double-precision performance.
Also includes Intel 300-series Optane persistent memory (2 terabytes per node), DDR5 memory (128 gigabytes per node), NVMe SSD storage (2 x 3.2 terabytes per node), and Nvidia NDR200 InfiniBand networking. The parallel file system provided by DDN provides 7.1 petabytes of 40 Gbps memory.
“The new supercomputer Pegasus is one of the first systems in the world to introduce 4th Gen Intel Xeon Scalable processors (formerly codenamed Sapphire Rapids), Intel Optane persistent memory (codenamed Crow Pass), and the Nvidia H100 Tensor Core GPU with 51 teraflops of breakthrough acceleration,”
reported the University of Tsukuba’s Center for Computational Sciences.
The project team reports that Linpack’s score for Pegasus is 3.47 petaflops, which should secure it a spot on the upcoming – in May – Top500 list. An increase in power efficiency is expected, which is significantly due to the Hopper GPU and persistent memory. Boku said it expects the Pegasus to be more energy efficient than Henri, the US-based H100-powered system that achieved the highest green ranking in November, achieving 65.09 gigaflops per watt.
According to University of Tsukuba measurements, Pegasus also has a higher Linpack efficiency, i.e. the usable fraction of theoretical peak flops: 54% for Pegasus versus 37.6% for Henri. Both metrics fall short of the ~65% average for the list. However, further optimizations may be forthcoming for both systems, so these figures are in some sense provisional until the next Top500 list is published.
Pegasus, which in the planning stages was named Cygnus-BD, will enable much larger simulations of traditional HPC applications in areas such as astrophysics, climate and biological sciences, and the large memory will also be used for big data and artificial intelligence workloads in a range of areas including drug discovery. Preliminary tests show that an astrophysics simulation code called ARGOT runs 1.86 times faster on Pegasus’ H100 GPU compared to Cygnus’ V100.