Perlmutter is the world’s most powerful computer dedicated to artificial intelligence. Launched at Berkeley Lab in the United States, this supercomputer will be used for scientific research. Inside? Hewlett Packard Enterprise engineering, AMD processors and 6,000 NVIDIA A100 GPUs.
With the collaboration of different technology brands, the laboratory has the second most powerful supercomputer in the world, second only to the Japanese Fugaku. It is, however, the most powerful when it comes to processing artificial intelligence data.
According to NVIDIA in a statement, the supercomputer is capable of reaching 180 PetaFLOPS (the Fugaku reaches 442 PetaFLOPS), making it the second most powerful supercomputer in the world. In addition, under extra performance conditions, it can reach four ExaFLOPS in AI processing.
Inside the computer are the high-end 64-core EPYC 7763 processors from AMD, plus the aforementioned 6,000 A100 GPUs from NVIDIA. HPE’s Cray Shasta architecture has been used to bring it all together. There’s one interesting detail, though: it starts up in two phases.
The first phase has 1,536 nodes. Inside each is a 64-core AMD EPYC 7763 CPU with 256 GB of DDR4 SDRAM and four 40 GB NVIDIA A100 GPUs. In this first phase the supercomputer achieves 60 FP64 PetaFLOPS of performance.
Once the second phase is implemented by the end of the year, 3,072 extra nodes will be added containing AMD EPYC 7763 CPUs and 512 GB of memory per node. With the second phase being implemented, 120 PetaFLOPS of performance is expected to be reached. The combined performance of the entire supercomputer will be 180 PetaFLOPS of performance and up to four ExaFLOPS of FP16 performance for AI applications.
Perlmutter is expected to be the fastest system on the planet when processing AI-related workloads. In one of its first projects it will help build the largest 3D map of the Universe visible to date.
In addition to the Universe mapping project, researchers will use the supercomputer to study subatomic interactions. A supercomputer like this one can accurately simulate the behavior of atoms over periods somewhat longer than the nanoseconds now achieved.