In the last year, the world of supercomputing has seen significant upheavals, with two innovative systems joining the Top500 list, surpassing the former frontrunner, Japan's Fugaku supercomputer.
Currently, the Aurora supercomputer, housed at the Argonne National Laboratory in the US, holds the second spot in the November 2023 rankings. This formidable machine, powered by Intel processors and accelerators, features 21,248 Xeon CPU Max 9470 52C 2.4GHz processors and 63,744 Intel Data Center GPU Max accelerators. However, due to incomplete performance measurements on all installed equipment, Aurora has not yet outperformed the top-ranked Frontier supercomputer at the Oak Ridge National Laboratory, also in the US. It's anticipated that within six months, the technical issues will be resolved, potentially propelling Aurora to nearly 2 EFlops, possibly making it the leader by June 2024. The Aurora's infrastructure components, such as interconnect, power supply, and cooling, are provided by Hewlett Packard Enterprise, similar to the Frontier system.
The Eagle system, engineered by Microsoft, holds the third position. This system is unique, being constructed entirely from standard modules typically used in Microsoft's data centers for the Azure cloud service. Remarkably, Eagle was assembled in just six months, a quarter of the time usually required for top-tier supercomputers. It employs 3,600 Xeon Platinum 8480C 48C 2GHz processors and 14,400 NVIDIA H100 accelerators, leveraging Microsoft Azure's standard infrastructure, including the Infiniband Interconnect (NVIDIA Infiniband NDR).
These leading systems exemplify the varied technological strategies employed in developing exascale (over 1018 Flops) supercomputers.
A noticeable trend in the Top500 ranking is the increasing disparity between the most and least powerful systems. The top ninety supercomputers contribute 80% of the total computational power of the list. The top system is now 500 times more powerful than the one at the bottom of the list, and this disparity is widening. However, the post-2016 crisis in supercomputing seems to have abated, with top-tier systems expected to grow in power by 1,000 times in 17-18 years, compared to the 23-24 years projected during the crisis period. The growth outlook for lower-tier systems in the Top500 is less optimistic, with a thousandfold increase expected over 27-28 years, reinforcing the notion that "the rich get richer while the poor get poorer."
In 2023, Russia's supercomputing capacity fell further behind global leaders like the US, the EU, China, and Japan, lagging by about 11.5 years compared to US capabilities. The technological gap, particularly in creating and possessing top-tier supercomputing systems, has widened to 11 years. This is a significant increase from the 2.5-year gap reported in 2012.
Globally, the supercomputing sector is evolving at a swift pace. In leading nations, key solutions to the technological challenges faced in the last ten years have been identified, encompassing both the hardware and software aspects of exaflops-scale supercomputers. The triumphs of three American systems (Frontier, Aurora, Eagle) and Japan's Fugaku are the result of robust governmental and international backing, the presence of multiple innovative (and competitive) strategies, and the integration of cutting-edge technological advancements across diverse fields.
