Frontier is two to three times more powerful than Fugaku, the supercomputer that held the top spot in the rankings for 2 years.
It’s quite a remarkable event in the very small circle of high-performance computing, or HPC: the Fugaku, a Japanese supercomputer that had been head and shoulders in its category for two long years – an eternity in this field – has ended up giving up the lead. Say hello to Frontier, the undisputed and indisputable new monarch of the elite supercomputer realm.
The machine is based at Oak Ridge National Laboratory in Tennessee. And to snatch the first place of the Top 500, a monthly ranking of supercomputers, the engineers had to pull out all the stops with high-end hardware. It is also interesting to note that unlike a large part of the machines in this category, Frontier has chosen to do without the Nvidia HPC GPUs of the Volta series.
monopoly of the heart
Instead of these devilishly powerful graphic processors which are part of the references on this segment, Frontier turned to the competition; it sails entirely under the AMD flag. According to PC Mag, the device works with 9,400 64-core EPYC processors et 37,000 AMD Instinct 250X graphics processors. In total, this represents a trifle of 8,730,112 processing subunits… A staggering amount of material that takes up no less than 74 standard cabinetsand which requires more than 22,000 liters of water to be cooled properly.
And as you might expect, this titan of silicon made short work of benchmarks for HPC platforms. As a reminder, these tests reason in terms of flops (Floating Point Operations Per Second). It is a unit of measurement that quantifies the number of operations that a system is able to perform in one second.
The “true” exascale era is finally here
Currently, there are already devices that are theoretically capable of approaching the exaflops course, i.e. a billion billion operations per second. This is for example the case of the Fugaku, which is indeed a computer called “exascale” On paper. But in real conditions, it has to settle for 442 petaflops. A figure that remains very impressive, so much so that it was a benchmark until today.
But the American machine has just taken high-performance computing into a new dimension; its designers claim an average score of 1.102 exaflops peaking at 1.685 exaflops; a figure three times higher than the record of 0.537 exaflops established by Fugaku! Note that the precise figures are to be taken with a grain of salt due to slight differences in the benchmarks, but the conclusion remains the same: Frontier is the very first supercomputer in the world to exceed the symbolic milestone of exaflops in real conditions.
More powerful, but also more profitable
The other interesting point is that this new version of Frontier is not only distinguished by its raw power; surprisingly, it also works wonders in terms of energy efficiency. It is measured in GFlops/watt, ie the number of operations per second that the machine can perform with an energy budget of one watt.
On the Green500, a classification of supercomputers according to their energy performance, Frontier moved directly into 2nd position and with 52 GFlops/watt. And the first is none other than… the Frontier Test Development System, the prototype that served as the basis for building this supercomputer. This is very surprising, since the heights of this ranking are usually occupied by devices that are significantly less powerful; the famous Fugaku, for example, does not even appear in the top 500 according to this criterion.
With this new explosion of scores at the top of the Top 500, the era of “exascale” computing is therefore becoming a reality. And given the speed at which these figures are progressing, we can expect that this famous exaflops will become the norm for supercomputers in the relatively near future.