Last November, the University significantly contributed to researchers nationwide after releasing DeltaAI, an artificial intelligence resource by the National Center for Supercomputing Applications.

DeltaAI is similar to ChatGPT but is more powerful and offers advanced computing resources to AI researchers, allowing them to progress through their research with faster and more accurate computational tools.

“What these systems are for is to provide computing resources to the nation’s researchers, so DeltaAI is providing computing to researchers all across the country,” said William Gropp, professor in Engineering and director of NCSA.

Funding, description of AI system

DeltaAI is funded by the National Science Foundation, which provides opportunities for the University to compete against other institutions and win funding for new systems. The University received around 10% of its funding from the NSF for hardware and operating on it, with other contributors, such as Illinois Computes, having a large part in funding DeltaAI.

“You’ll see some inconsistencies in the numbers because of the way the National Science Foundation funded it; there was an initial award which was $10 million for the hardware and then an additional $12 million over five years for operations,” Gropp said. “But then we got several additional supplements, and there was some additional funding to operate it, and then an additional $5 million to increase the size of the hardware.”

Unlike normal computers, the systems by the NCSA are much more powerful and make up an advanced network for DeltaAI, allowing researchers to input data into the AI resource. 

“The way these computers are built is that they’re really made up of a number of computing nodes,” Gropp said. “You can sort of think of a computing node as a really amped-up gaming laptop, and so we have just over 100 in DeltaAI. Because they’re really amped-up nodes, we can actually bring them together, so DeltaAI actually shares a high-performance network.”

Specifics on AI’s computing power

DeltaAI is much faster and more powerful than the University’s reigning supercomputer, Blue Waters. According to the NCSA’s website, Blue Waters is about 3 million times faster than your average laptop and can perform 13 quadrillion calculations per second.

DeltaAI is much more powerful because it implements state-of-the-art graphics processing units from Nvidia, which designs high-quality hardware and competes with companies like Google and Amazon to create better processors. 

“One of the things about the Nvidia GPUs is that we have H100s in DeltaAI that have a more general purpose than those specialized processors,” Gropp said. “So what Nvidia has done is taken a GPU, which is itself somewhat specialized, but it’s also good at large classes of computations, and they have added additional features to make them especially good.”

The same H100 GPUs DeltaAI uses were also used by Elon Musk to develop an AI supercomputer in Memphis. 

Challenges for AI

One of the challenges faced by AI researchers is in researching explainable AI; one works with large data sets, which requires a ton of computer output. DeltaAI alleviates this problem to some extent with powerful hardware.

“Just training one of those large language models is an enormous computational task,” Gropp said. “Even once you have a trained model, running it still requires a pretty hefty computing environment. If you’re conducting this sort of research and you need to be able to run these models and explore them, you need a resource like DeltaAI. It means that the researchers will be able to get access to state-of-the-art GPUs and data systems to allow them to do that research.”

A more practical example of the necessity for higher-power resources for research is the National Deep Inference Fabric, a collaborative project between the University and researchers at Northeastern University. It partners with the NCSA to use DeltaAI to provide researchers with optimal computing capabilities for their research projects.

“The idea behind that is to provide an environment in which researchers would be able to explore how the results come out of these models,” Gropp said. “We are trying to understand why what comes out … What we proposed (to) the National Science Foundation was that AI researchers need resources, and so we focused on the AI part for this.”

Metrics for AI

The metrics indicating the success of DeltaAI relate to the number of papers published and the impact of those papers. NCSA will also measure the number of users on it, the number of projects and the demand for the resource.

Computer operations

The NCSA assesses the computing output of DeltaAI by measuring the number of arithmetic operations, such as addition and multiplication, that the machine can perform in one second. A computer uses floating-point numbers, which are integer numbers that the computer stores in very small units of data or bits.

“A laptop is amazing,” Gropp said. “They can do about a gigaflop. That’s about a billion, or 10 to the ninth operations per second. The fastest machines in the world are now a bit over an exaflop. An exaflop is 1,000 petaflops, a petaflop is 1,000 teraflops, a teraflop is 1,000 gigaflops. So giga is (10 to the) ninth, tera (to the) 12th, so 10 to the 15th operations per second is a petaflop.”

Regarding storage, AI research can work with numbers represented by much fewer bits. This means a relatively small number of digits was used, causing the range of the exponent to get narrower and affecting the precision of the research. This can be changed by implementing half-precision or double-precision numbers. Half-precision numbers use less data but have less precision, while double-precision numbers are more accurate but require more data. 

“When you’re working with a trained model, it turns out that you don’t need much precision,” Gropp said. “You can work with 16-bit numbers, and that’s two bytes. So you look at how much data you’re storing: A classic double-precision number uses eight bytes, and a half-precision number only uses two. So it only takes one-fourth of space, and if your hardware is designed specially for it, like the Nvidia GPUs or special processors that Amazon and Google are developing, then you can process those smaller items really fast.”

DeltaAI computation

DeltaAI computes at 20 petaflops for double precision, 40 for single precision and 80 for half-precision. This is because performance is inversely proportional, where if it is half the number of bits, the performance is doubled. It is slightly different for AI, where the shorter precision numbers are much faster, allowing for more operations per second.

“(Twenty) petaflops in double precision is still faster than Blue Waters, with 600 petaflops for the half-precision, the full machine with the supplement is more like 30 petaflops of double precision and over 900 petaflops for the half-precision, and so that’s more relevant at looking at AI work,” Gropp said. “Those are sort of raw peak numbers; those are guaranteed not to exceed numbers. Actual applications will be a bit slower. Now, we will be looking at some of the benchmarks both in computation and in AI, but in the end, what we’re really looking at is how this accelerates the nation’s science.”

Future for DeltaAI

The hope for DeltaAI is to apply the resource to 22 different areas of AI research. Future developments of the resource include making models more trustworthy and reliable and making them faster and able to conserve more energy.

With DeltaAI’s computing power, better mathematical models could be used at a much larger scale in fields of physics, such as fluid flow modeling. Instead of using an approximate model, good models can be used to model fluid flow as if it is continuous since the AI can use large sets of data. 

“The basic idea is that there are lots of problems where we can use AI to gain insights out of the data and use that to help us solve problems in other areas of scholarship,” Gropp said. “Whether it’s in engineering or science or beyond that as well, and with DeltaAI, we want to be able to enable researchers to better understand how to take advantage of AI as a tool to make their research advance faster and let them gain insights that they have not been able to gain before.”

[email protected]