In 2018, the Summit supercomputer was installed at Oak Ridge National Laboratory, with a peak computing power of 200 petaflops, 200,000 trillion floating-point operations per second at the time. The appearance of the Summit supercomputer displaced Chinese computer Sunway TaihuLight from the first position of Top500, and much later the Summit supercomputer gave way to the Japanese Fugaku system, remaining at the second position at the moment.
In just the last four years, the complexity of the tasks performed by the most powerful supercomputers has grown so much that the performance of the Summit system is already catastrophically insufficient for these tasks. "The Summit system can now provide only one-fourth or one-fifth of the required power," says Justin Whitt, head of supercomputing at Oak Ridge, "This significantly limits the number of research projects in which the Summit supercomputer can be used."
The obvious solution to the problem is to build a new and even more powerful supercomputer, and that is exactly what is happening now at Oak Ridge Laboratory. The new system, now in the build stage, is called Frontier, and when it is fully operational, it will have a peak estimated power of just over 1.5 exaflops.
Of particular note, the Frontier system will be seven times more powerful than the Summit system, but the amount of power it will consume is only double that of the Summit system. It is estimated that, at maximum load, Frontier will consume 29 MW, comparable to the power consumption of a small city such as Cupertino, California.
This increase in computing efficiency is due to the placement of more computer hardware in a smaller space. Each Frontier supercomputer rack will be so densely packed that its weight will be comparable to that of an average truck. The use of AMD EPYC processors focused on artificial intelligence and Radeon Instinct GPU accelerators optimised specifically for exascale-level computations will also play a significant role in boosting performance and efficiency.
According to initial plans, an even more powerful supercomputer called Aurora was to be commissioned at Argonne National Laboratory in 2022. But its commissioning has been delayed due to a disruption in the supply chain of the necessary components by Intel, which, due to the crisis in the semiconductor industry, is simply not able to produce the required number of CPUs and GPUs by the deadline.
Back to the Frontier supercomputer. Oak Ridge Labs predicts that the machine will be fully assembled by the end of this year and will be operational during 2022.
But will the Frontier system make any difference to the Top500, where Japan's Fugaku is currently number one, with 2 exaflops of performance on a particular type of test? All of the systems in the ranking are tested on 64-bit floating-point calculations, which are required to solve three-dimensional systems of partial differential equations used in various types of physical simulations.
But the main type of tasks that the new Frontier system will perform is the training and use of artificial neural networks, for which 16-bit number digits are also sufficient. And the Frontier system, which has been optimized for this purpose, is not expected to be able to perform calculations with that precision.
