Los Alamos National Laboratory, on September 11, 2023, announced a groundbreaking achievement in the field of machine learning. Their team has successfully developed an algorithm capable of processing data larger than a computer's memory. This algorithm, which sets a new world record for processing massive data sets, was put to the test on Oak Ridge National Laboratory's Summit supercomputer.
During the test, the algorithm showcased its capabilities by successfully processing a dense matrix of 340 terabytes and a sparse matrix of 11 exabytes. This impressive feat was achieved using 25,000 GPUs. What makes this algorithm particularly remarkable is its scalability, as it demonstrated equal efficiency whether run on regular laptops or advanced supercomputers. This advancement is a significant breakthrough that addresses critical hardware limitations in processing data-intensive applications across various domains, such as cancer research, satellite imagery analysis, and national security science.
Ismael Boureima, a computational physicist at Los Alamos National Laboratory, explained that the algorithm breaks down large data sets into smaller units that can be processed with the available resources. This "out-of-memory" implementation of the non-negative matrix factorization method allows for the factorization of larger data sets than previously possible on a given hardware. Boureima emphasized that this algorithm is a useful tool for keeping up with the exponentially growing data sets that researchers and scientists are now faced with.
The innovative machine-learning algorithm developed by Los Alamos National Laboratory leverages hardware capabilities, such as GPUs, to enhance computational speed and facilitate rapid data transfer between computers. It is highly efficient in multitasking and can handle multiple tasks concurrently. This achievement is part of the broader SmartTensors project at Los Alamos, which aims to develop a repertoire of high-performance algorithms. Boureima also highlighted that while scaling to 25,000 GPUs is impressive, the algorithm is still useful on desktop computers, allowing for the processing of data that was previously unattainable.
