The US Department of Energy has yet to launch Frontier — on track to become the country’s first exascale supercomputer — but scientists are already getting a taste of what the machine will do, thanks to a powerful early version of it called Crusher.
Crusher is in fact a small part of the Frontier system. Frontier will eventually feature over 100 Cray Shasta cabinets, with a performance of 1.5 exaflops. Crusher, meanwhile, is a 1.5 cast iteration of Frontier. It features 192 nodes connected by HPE’s Slingshot Interconnect, with each node containing one optimized 3rd Gen AMD EPYC processor and four AMD Instinct MI250X accelerators.
Crusher occupies only 44 square meters of floor space. It’s 1/100 the size of the DoE’s 27-petaflop Cray XK7 Titan supercomputer, which was decommissioned in 2019, but it’s faster than the entire 4,352-square-foot system.
Currently, four renowned projects at the Oak Ridge Leadership Computing Facility (OLCF) are executing scientific codes on Frontier’s architecture through Crusher. The results so far are impressive.
For example, the CANcer Distributed Learning Environment (CANDLE) project successfully ran a Transformer model on Crusher, achieving an 80% speed improvement over previous systems. The project develops next-generation natural language processing models for precision medicine using “Transformers”, which can identify invisible links between words in clinical text. Developed from a collaboration between DOE and the National Cancer Institute (NCI), the project aims to provide the NCI with better, more accurate models for cancer surveillance.
Meanwhile, a nuclear physics code that can run massive simulations of nuclei sees 8-fold accelerations on Crusher, compared to the DoE’s current Summit supercomputer. And one of the first astrophysics codes to be rewritten for Frontier is Crusher going 15 times faster than the base Fall 2019 tests on the Summit supercomputer.
The Frontier was expected to become the fastest supercomputer in the world when it arrived at Oak Ridge National Laboratory last year, but it is still undergoing integration and testing. The Frontier system is expected to be fully operational for users by 2023.