Lincoln Labs Launches Most Powerful AI Supercomputer at any university in the United States | MIT News

Lincoln Laboratory Supercomputing Center (LLSC)’s new TX Generative AI Next (TX-GAIN) computing system is the most powerful AI supercomputer at any American university. TX-GAIN recently released the latest ranking of the Top 500, which lists two top supercomputers, and TX-GAIN joins the ranking of LLSC’s other powerful systems, all supporting research and development at Lincoln Labs and MIT campuses.

“TX Gain will enable our researchers to achieve scientific and engineering breakthroughs. The system will play an important role in supporting generated AI, physical simulations and data analysis in all areas of research,” said Jeremy Kepner, a researcher at Lincoln Laboratory.

LLSC is a key resource to accelerate innovation in Lincoln Labs. Thousands of researchers use LLSC to analyze data on federally funded research projects, train models and operational simulations. For example, supercomputers have been used to simulate the encounters of billions of aircraft to develop collision avoidance systems for the Federal Aviation Administration and to train models in the Department of Defense’s autonomous navigation complex mission. Over the years, the capabilities of LLSC have been crucial to numerous award-winning technologies, including technologies that improve airline safety, prevent the spread of new diseases and contribute to the hurricane response.

As the name implies, TX Goain is specifically used to develop and apply the generation of AI. Traditional AI focuses on classification tasks, such as determining whether a photo depicts a dog or a cat, and the generated AI produces entirely new output. Kepner describes this as a mathematical combination of interpolation (filling in blanks between known data points) and extrapolation (expanding data beyond known points). Today, generated AI is known for using large language models to create human-like responses to user prompts.

At Lincoln Labs, the team will generate AI to apply it to various fields beyond large language models. For example, they are using the technology to evaluate radar signatures, supplement weather data for coverage lacking coverage, root network traffic anomalies, and explore chemical interactions to design new drugs and materials.

To enable this intense computing, in addition to traditional high-performance computing hardware, the TX GAIN is powered by over 600 NVIDIA graphics processing unit accelerators designed specifically for AI operations. TX GAIN has a peak performance of two AI Exaflops (two trillion floating point operations per second), making it the top AI system in the university and the Northeast. Researchers have noticed since TX Goain went live this summer.

“TX Goain can not only model protein interactions more than ever, but also model protein interactions with more atoms. This new computing power is a game changer in the protein characterization efforts in biological defense,” said Rafael Jaimes, a researcher at Lincoln Laboratory.

LLSC’s focus on interactive supercomputing makes it particularly useful to researchers. For years, LLSC has pioneering software that allows users to access their powerful systems without being an expert in configuring algorithms for parallel processing.

“LLSC has been trying to make supercomputing like working on a laptop,” Kepner said. “The complexity of requiring competitive data and analytics today goes far beyond what you can do on a laptop. But with our user-friendly approach, people can run models and get answers quickly from the workspace.”

In addition to providing support programs only at Lincoln Labs, Texas GAIN is also strengthening research collaborations with the MIT campus. Such collaborations include the Haystack Observatory, the Center for Quantum Engineering, Beaver Works and the Air Force Division MIT AI Accelerator. The latter initiative is rapidly prototyping, scaling and applying AI technology to optimize flight schedules for global operations as an example of a lab.

The LLSC system is located in energy-efficient data centers and facilities in Holyoke, Massachusetts. LLSC researchers can also meet the huge energy needs of AI and lead research on various stress relief methods. A software tool they developed can reduce the energy of training AI models by up to 80%.

“LLSC provides the functionality needed to conduct leading research, while simultaneously being cost-effective and energy-efficient,” Kepner said.

All supercomputers on LLSC are named after the 1956 Lincoln Laboratory Transistor Experiment Computer Zero (TX-0). With TX GAIN, LLSC continues this legacy.