Saudi Arabia’s KAUST Collaborates with Cerebras Systems for Seismic Processing Breakthroughs
Last Updated on November 9, 2023 by Ameer Hamza
KAUST, the top research university in Saudi Arabia, has partnered with Cerebras Systems, a pioneering US-based company in generative AI acceleration. Their collaborative efforts are focused on seismic processing and leveraging the powerful Condor Galaxy AI supercomputer. Cerebras, in partnership with G42 from Abu Dhabi, built this cutting-edge machine, earning them a spot as finalists for the renowned 2023 Gordon Bell Prize.
Gordon Bell pioneered computer science and high-performance computing, and the Gordon Bell Prize is a major honor named in his honor. The award, now in its 35th year, is given each year to honor excellent accomplishments in the field of high-performance computing (HPC).
Breaking Boundaries: How KAUST and Cerebras Redefined Seismic Processing with AI
The research conducted by KAUST and Cerebras focuses on multi-dimensional seismic processing. This study area is important for understanding the Earth’s resources and leading the world toward a sustainable, low-carbon future. Through this creative cooperation, Cerebras Systems’ cutting-edge technology was able to interpret seismic data at previously unattainable speeds and accuracy levels.
The key to this achievement is the development of a specialized kernel called Tile Low-Rank Matrix-Vector Multiplication (TLR-MVM). This kernel is specifically designed to take advantage of the advanced architecture found in Cerebral CS-2 systems used in the Condor Galaxy AI supercomputer. The outcome was a seismic processing solution that achieved production-grade accuracy. It set a new record for sustained memory bandwidth, reaching an impressive 92.58 petabytes per second (PB/s). This is noteworthy because bandwidth is often the limiting factor in modern processor-rich architectures. The achievement shows how important AI-customized systems can be in bringing about a new era of seismic algorithms that are much better than before.
Andrew Feldman, who is both a co-founder and the CEO of Cerebras Systems, expressed his pride in the groundbreaking partnership with KAUST researchers and their use of the Condor Galaxy AI supercomputer. Feldman also said that he was very grateful to Gordon Bell for recognizing their outstanding work in pushing the limits of multi-dimensional seismic processing. This project could lead to big changes in many areas, like climate and weather modeling, computational astronomy, wireless communication, seismic imaging, and more.
Extreme Computing Research Center (ECRC) Principal Research Scientist and lead author Hatem Ltaief expressed his excitement at the study’s success. He pointed out that disaggregated memory requires complex algorithmic developments and that it had been a goal for a long time to work with Cerebras developers to apply these developments and use the hardware to its fullest potential.
David Keyes, director of the ECRC, also expressed enthusiasm over the adaptability of wafer-scale technology, adding that it goes beyond its initial use in neural network training optimization. Such versatility, he said, is a notable addition to the history of architectural crossovers that have won the Gordon Bell Prize.
The work done by KAUST and Cerebras on Condor Galaxy -1, the AI supercomputer created by Cerebras and G42, illustrates the revolutionary potential of Cerebras CS-2 as a substitute that can deliver record-breaking performance for formerly memory-bound applications.