EN
Today, the Oak Ridge National Laboratory (ORNL) has deployed SambaNova Suite to expand its research with secure, trustworthy, and energy-efficient AI. The lab is tackling problems of national importance and will become the latest US national lab to deploy SambaNova Suite, which combines DataScale hardware, SambaStudio software, and the unique Composition of Experts (CoE) architecture for running multiple models.
ORNL is home to Frontier, the most powerful supercomputer in the world, and the lab uses its vast computing resources to conduct world-changing research in areas that include energy and security, medicine, and climate science. SambaNova will collaborate with the lab to create an AI Center of Excellence to elevate the use of AI in those efforts. ORNL will deploy two DataScale nodes, each powered by 16 SambaNova SN40L chips.
“Our research is focused on secure, trustworthy, and energy-efficient AI. We hope to learn and use the capabilities of SambaNova’s systems to enhance what we are doing. SambaNova’s fast inference, energy efficiency, and its Composition of Experts architecture will enhance the capabilities in our AI for science portfolio,” said Prasanna Balaprakash, Director of AI Programs at ORNL. SambaNova’s platform will allow multiple models to be run and queried in parallel so their answers can be combined to make predictions better. “It’s a transformative capability for the kind of problems we work on across multiple domains.”
ORNL is developing its own specialized AI models, Balaprakash said, and the lab will run sets of similar tasks across multiple models trained on different kinds of data — say, a model trained on scientific texts and another on scientific images — so that many predictions can be combined into an aggregated result, leading to better predictions overall. ORNL scientists conduct research in several areas, including material science, climate modeling, neutrons, as well as nuclear fusion and fission. SambaNova’s platform will be used for parallel inferencing across models trained on data from all of these fields. In one case, the platform will be used for inference on a foundational climate model that was initially trained on Frontier, Balaprakash said.
The first steps will involve evaluating many models to see how they behave under different conditions and to measure their trustworthiness. Eventually, ORNL scientists will be able to query many different models across multiple domains of science for correlation. For instance, scientists working on potential designs for a stable fusion reactor will be able to query models trained on different branches of materials science in order to look for connections they wouldn’t have otherwise found. This ability to query many models at once across different fields of science will be a powerful catalyst for the work ORNL does.
Another key factor in the decision was SambaNova’s capabilities for fast inference at scale with greatly improved energy efficiency. As ORNL evaluates different models, the SambaNova platform’s fast inference capabilities can be scaled up to run reasoning tasks while using less energy than would be required on Frontier. “It’s much more efficient to do the inference on a platform that is specialized and customized for faster inference,” Balaprakash says. “If we can do this with a fraction of the energy costs, that’s a big win.”
Overall, Balaprakash said he expects the deployment of SambaNova Suite to help increase ORNL’s productivity and lead to more scientific discoveries. “The ability to extract information across multiple domains is very, very exciting. It could be a big boost for the science we do at ORNL.” Learn more about SambaNova Suite.
