Academia’s Fastest Supercomputer Delivers 700 PFLOPS of AI Performance

Article By : Brian Santo

The University of Florida is building a supercomputer built around Nvidia GPUs using an architecture that helped vault Nvidia's own supercomputer into seventh place on the list of most powerful machines in the world...

The University of Florida will soon have the most powerful supercomputer in academia, a machine that will be powered by Nvidia GPUs. The supercomputer, designated the HiPerGator 3.0, will be the centerpiece of UF’s newfound ambition to become one of the leading universities in the world for artificial intelligence (AI) research.

All of this recently became possible when Florida alumnus and Nvidia co-founder Chris Malachowsky decided to donate $25 million to the school. As a company, Nvidia vowed to donate $25 million in hardware, software, training and services. The university is also kicking in $20 million to support its AI programs.

The new high-performance Gator 3.0 (named after the university’s mascot) is based on Nvidia’s new DGX SuperPOD architecture. This is the same architecture that Nvidia’s own Selene supercomputer was built on. Selene debuted at No. 7 on the Top 500 list of the world’s most powerful supercomputers in June; it is also one of the most energy-efficient supercomputers in existence. Nvidia said it built Selene and had it up and running in just a few short weeks.

HiPerGator will feature 140 DGX A100 systems, with 700 petaflops of peak AI performance, 1,120 Nvidia A100 GPUs, 170 Mellanox Quantum 200G InfiniBand switches, 15 kilometers of optical cable and 3.5 TB of high-performance storage, according to an Nvidia spokesman.

“The initial two systems UF is using for preliminary testing and development will be augmented by the 140-node DGX SuperPOD being delivered later this year,” the spokesman added.

Asked about the performance expected of HiPerGator 3.0, the spokesman said “a DGX SuperPOD of 140 DGX A100 systems and Nvidia Mellanox InfiniBand networking delivers 700 petaflops of AI performance. There are 1,120 GPUs in 140 DGX servers, with each GPU capable of delivering 624 TFLOPS of AI compute (mixed precision), which yields the 700 petaFLOPS figure.”

A rendering of the HiPerGator 3.0 supercomputer at the University of Florida

The press conference announcing the supercomputer and UF’s AI program featured Malachowsky, Nvidia CEO Jensen Huang, multiple university officials, and even Florida Gov. Ron DeSantis. All spoke of how the company, the university, and the state have specific plans to collaborate on AI research and adopting AI technology. One speaker noted how the state of Florida has a database of anonymized medical data of 14 million people that can be used for research.

Specific elements of the AI program include collaboration with the NVIDIA Deep Learning Institute which will work with UF on developing new curriculum and coursework, and the the establishment of a new Nvidia AI Technology Center at UF. The university is also creating an Equitable AI program, which aims to convene faculty members across the university to create standards and certifications in developing tools and solutions that are cognizant of bias, unethical practice and legal and moral issues in AI.

Explore More:

Take advantage of wide-ranging AI opportunities. GigaOm report shows you how to devise, define, and deploy the right AI for job.

Leave a comment