Lawrence Berkeley National Laboratory (Berkeley Lab) of the National Energy Research Scientific Computing Center (NERSC) announced that the new supercomputer Perlmutter has officially launched. This HPE Cray EX supercomputer is named Perlmutter in memory of the laboratory astrophysicist Saul Perlmutter, who won the 2011 Nobel Prize in Physics.
Perlmutter is built on the HPE Cray Shasta platform. It is a heterogeneous architecture system that includes CPU and GPU acceleration nodes. Its performance is three to four times that of the supercomputing Cori currently used by the National Energy Research Scientific Computing Center. The system will be delivered in two phases as planned.
The first stage is to start installation in early 2021. There will be 12 GPU acceleration cabinets, which can accommodate 1536 nodes, 35PB of all-flash storage, and 5TB/s data transfer speed. Each node has an AMD EPYC 7763 processor, as well as 256GB of memory and four Nvidia A100 40GB GPUs connected via NVLink.
During this period, Perlmutter can provide FP64 performance of 60 PetaFLOPS or FP16 performance of 3.823 ExaFLOPS. This performance is enough to make Perlmutter enter the top ten position of the world’s supercomputer system, but Perlmutter will not stop there.
The second phase will start later in 2021. 12 CPU cabinets will be added and pure CPU nodes will be used. There are 3072 nodes in total, all using AMD EPYC 7763 processors. Each node has 512GB of memory, and its FP64 performance reaches about 120 PetaFLOPS. After two phases of deployment, Perlmutter’s FP64 performance can reach 180 PetaFLOPS, which will surpass the world’s second-ranked supercomputing Summit, but still lags behind Fugaku, the performance of which has reached 442 PetaFLOPS.