TL;DR: Private companies urgently need additional GPU computing capacity to train their new generative AI services, but face significant challenges in obtaining it. In contrast, researchers involved in US supercomputing projects currently have access to powerful Nvidia GPU nodes at a significant discount, albeit for a limited time.
Users working with the National Energy Research Scientific Computing Center (NERSC) and the U.S. Department of Energy can now run their tasks on Perlmutter’s GPU nodes at half the price of what they would have paid just a few weeks ago . This special offer is available until the end of September. According to NERSC’s Rebecca Hartman-Baker, this represents an excellent opportunity to begin computationally intensive research after the summer break.
Mother of pearl serves as NERSC’s primary high-performance computing system and consists of a supercomputer equipped with 3,072 AMD “Milan” EPYC CPUs and 1,792 NVIDIA A100 GPU-accelerated nodes. This system includes several technological innovations aimed at accelerating the scientific productivity of researchers accessing NERSC services.
Importantly, it is intended exclusively for scientific research purposes. The name “Perlmutter” commemorates Saul Perlmutter, the distinguished American astrophysicist who led the team that won the Nobel Prize for their groundbreaking evidence confirming the accelerating expansion of the universe.
NERSC’s 50% discount offer is currently incentivizing the execution of scientific orders on Perlmutter GPU nodes, circumventing the typical “year-end crisis” characterized by longer query times and delayed order fulfillment. Hartman Baker conveyed announced in an email announcement that the use of the HPC system is currently benefiting the entire NERSC community by helping to distribute computing demand more evenly throughout the year.
Any order or portion thereof running from September 6th to early October 1st will incur only half of the standard fees. Thanks to this time-limited discount rate, Hartman-Baker explained that a three-hour job on seven GPU nodes would only cost 10.5 GPU node hours, while without the discount the same job would have cost 21 GPU nodes. hours
NERSC also offers users additional support through the Perlmutter GPU.Virtual office hours“, which provides support for getting started with supercomputer GPU nodes and addresses concerns about insufficient allocation, poor performance, and more. Perlmutter’s total processing performance in phase one, which ended on May 27, 2022, reached 70.9 PFLOPS.
As pointed out by Glenn K. Lockwood, HPC storage expert at Microsoft, who was the first to do this report Due to NERSC’s special offer, the AI industry is currently grappling with a “GPU crisis” that is unlikely to be resolved anytime soon. If the Department of Energy were to lease its “idle” computing capacity for commercial workloads, the U.S. government could potentially generate a significant revenue stream.