Remember the world’s largest GPU that Huang Jen-xun pulled out of the oven two months ago? This time, Nvidia is coming with something even bigger!
For a computer, running speed is one of the most important performance criteria, in general, the faster the better. So just how fast is Nvidia’s AI supercomputer this time?
At Nvidia’s GTC 2020 presentation on May 14, Huang talked about the next generation of DGXSuper Pod clusters. At the time, he touted the DGXSuper Pod as consisting of 140 DGXA100 systems and capable of 700 Petaflops AI.
Today, the “fastest AI supercomputer in academia” is finally no longer an invisible number. Nvidia and the University of Florida (UE) have announced plans to build the fastest AI supercomputer by enhancing the capabilities of the existing Hipergator supercomputer using the DGXSuper POD architecture.
Nvidia has also made a commitment to have the 700 Petaflops system up and running in early 2021.
What good is super computing power?
You think AI is just “baby” technology that we see, can hold simple conversations, recognize faces, or do common tasks? If so, why did the world’s top scientists devote so much energy to research and development? They would have quit long ago. For example, what we have achieved so far is just the tip of the iceberg. Just like the brain development of ordinary people, there are still many talents that are not being utilized.
Researchers in the field of artificial intelligence are naturally taking the longer view, arguing that powerful computers combined with reinforcement learning and other techniques can enable advances that change the model of artificial intelligence. In human terms, more possibilities can be developed for artificial intelligence.
A recent paper published by researchers from MIT, the MIT-IBM Watson Artificial Intelligence Lab, Underwood International School, and the University of Brasilia found a “strong dependency” on deep learning improvements.
In 2018, researchers at OpenAI released an analysis showing that from 2012 to 2018, the number of computers used in the largest AI training increased by more than 300,000 times, doubling in 3.5 months — far faster than Moore’s Law.
Florida State University and Nvidia say the improved HiperGator will provide teachers and students with AI tools for use in key areas including rising sea levels, an aging population, data security, personalized medicine, urban transportation, and food security.
The AI model at UF Health has been deployed into a system called DeepSofa to help collect, organise and monitor patients’ conditions in real time.
HiperGator, which is about 18 times more powerful than the University of Texas at Austin’s Frontera, was one of the first companies to use Nvidia’s DGX A100 system. The DGX A100 contains eight A100 Tensor Core GPUs based on 7 nm amps, offering 320 GB of memory and the latest high-speed Mellanox HDR 200Gbps interconnection.
According to Nvidia, a single A100 GPU with 54 billion transistors can perform 5 petaflops.
As part of the push, Hipergator will receive 140 DGX A100 systems powered by 1,120 NVIDIA A100 Tensor Core GPUs, plus 4 PB bytes of storage space from the DNT and 15 km of fiber optic cable. It will also benefit from Nvidia’s suite of AI application frameworks, including data analytics, AI training and inference acceleration, and recommendation systems.
Comprehensive and deep cooperation in talent, technology and platform
In addition to Hipergator, Nvidia also said its partnership with UF will extend to ongoing support and collaboration in three major AI areas. The Nvidia Institute for Deep Learning will partner with the University of Florida to develop courses, courses, and programming to meet the needs and encourage the interests of young people and adolescents in science, technology, engineering, math (STEM) and artificial intelligence.
In addition, Nvidia will establish a Nvidia AI Technology Center at UF, where graduate researchers and Nvidia employees will collaborate to advance the field of artificial intelligence. Nvidia solution architects and product engineers will work with UF to install, operate and optimize supercomputing resources on campus, including Hipergator.
The University of Florida, one of Nvidia’s first partnerships, says the collaboration will lay the foundation for the integration of AI with all of its disciplines, making it a “ubiquitous part” of the institution’s academic endeavors.
In order to better promote the process of building a super AI computer project, the first thing to do is to recruit more talent to “help build the team”.
To attract more talented people in related fields, UE will also offer certificate and degree programs in artificial intelligence and data science, as well as course modules focused on technology and industry, and has pledged to recruit 100 more teachers focused on artificial intelligence.
With the foundation in place, the University of Florida plans to train 30,000 graduates with AI skills by 2030. Achiving that goal, of course, can’t just be done by our graduates. It can also involve partnering with historically black colleges, Hispanic-serving institutions and K-12 programs, as well as building fair artificial intelligence programs. The program will seek to have faculty across the school create standards and certifications to develop tools and solutions that recognize bias, unethical practices, and legal and ethical issues.
In addition, UF intends to work with other institutions and academic teams to conduct research and recruitment for UF’s graduate program, and to provide training in artificial intelligence.
The goal is great, but where will the money come from?
The goal is great, the plan is detailed, then the question is, the foundation of all this — where will the money come from?
In fact, both these initiatives and the Hipergator improvements so far have been funded by a $50 million grant from UF alumnus and Nvidia researcher Chris Malachowski. Another $25 million in hardware, software, training and services from Nvidia.
To keep the project going, UF will also invest more than $20 million to upgrade its campus data center around its new machines.
You have the money, the people, the technology, the platform, and everything. It’s going to be a 700 Petaflops AI system by the beginning of next year.
