American institute equips AI supercomputer with 6000 Nvidia A100 GPUs

The US National Energy Research Scientific Computing Center is commissioning Perlmutter, a supercomputer focused on artificial intelligence. The supercomputer contains more than six thousand Nvidia A100 accelerators and fifteen hundred AMD EPYC 7763 CPUs.

Placed in Berkeley Lab’s Nersc’s Shyh Wang Hall facility, Perlmutter is named after Saul Perlmutter, a physicist who co-won the Nobel Prize in 2011 for his research into the speed at which the universe is expanding. More than 7,000 scientists can use the high-performance system, which is three to four times as powerful as Nersc’s Cori supercomputer, which was the most powerful in the institute to date.

The cluster is based on HPE Cray’s Shasta platform and will be built in three phases. Phase one is ready and offers 1536 nodes, each with an AMD EPYC 7763 processor and four A100 GPUs connected via NVlink. The system also offers a Luster file system of 35 petabytes of flash memory. This gives scientists access to more than four exaflops of mixed-precision computing power.

They can use the computing power for, among other things, making a 3D representation of the visible universe, Nvidia describes. To do this, huge amounts of data from the Dark Energy Spectroscopic Instrument have to be processed, something that took weeks to months on the previous system, which should reduce Perlmutter to a few days.

Later this year, the supercomputer should be upgraded at phase two with 3072 CPU nodes, each with two AMD EPYC 7763 processors and 512GB of memory per node.

Leave a comment