The biggest event in the world of science and information technology that will mark 2021 is certainly the launch of the most powerful supercomputer on the planet - ORNL (Oak Ridge National Laboratory) "Frontier".
"Frontier" will enable performance that exceeds 1.5 ExaFLOPS. For comparison, currently, the fastest supercomputer in the world is the Japanese "Fugaku" with a theoretical maximum of 0.53 EFLOPS. This kind of computing power was created to break new boundaries in scientific research. The most difficult scientific problems require the greatest possible budgetary power. Exploring the possibilities of applying fusion energy, finding new materials, unlocking the secrets of the universe, all these and many other problems are key to the further development of humanity.
On the occasion of preparations for the introduction of full capacity, ORNL has created a special "CAAR" program in the selection of which there are eight major scientific problems on which "Frontier" will work as a kind of prologue for the transition to full performance mode. The most demanding of these eight problems is, of course, the precise simulation of the entire galaxy, which is similar in size to our Milky Way.
The "CHOLLA" program for simulating entire galaxies is the work of astrophysicists from the renowned Princeton University. This program has so far worked on a "Titan" supercomputer, whose power is 0.027 EFLOPS! With that in mind, it is not surprising why "CHOLLA" was chosen as the first program to be run on the new supercomputer.
An additional reason why "CHOLLA" was chosen to test "Frontier" is that it relies entirely on GPU processing, which can be concluded when this abbreviation is developed: "Computational Hydrodynamics on Parallel Architecture".
Because Titan used Nvidia Kepler GPUs, a team of Princeton developers, in collaboration with AMD and HPE (HPE bought Cray), allowed engineers to transfer the complete "CHOLLA" program code very easily and very quickly. from NVIDIA's CUDA to AMD's ROCm, using the HIP translation layer. According to Reuben Budiard, switching the code was very easy and did not require more than a couple of hours of work. According to the initial "CHOLLA" tests on AMD Instinct MI50 accelerators, using ROCm, it has equivalent performance as on NVIDIA Tesla V100 accelerators, using CUDA. Switching to AMD Instinct MI100 accelerators gives an instant acceleration of 40% without any change in the ROCm code itself. The team expects even greater acceleration with better optimization for the new AMD CDNA architecture.
Why is every “TFLOPS” important with this type of budget? The simple answer is due to the resolution of the details. Namely, "CHOLLA" has the ambition to simulate both the fluid dynamics of gases and the gravitational interactions between stars in gas nebulae through the simulation of the birth and death of each star in the galaxy, with the greatest possible precision. Without adequate computational power, this key aspect of stellar influence on the formation and shape of galaxies cannot be simulated at the required resolution and scale.
Thanks to the computing power of AMD Instinct MI100 accelerators, based on the latest "CDNA" architecture, the team from Princeton University will be able to run their "CHOLLA" program in full resolution, which will allow for the first time to compare simulated results inside supercomputers with visual data collected over the years. telescopic laboratories around the planet, as well as the extraterrestrial Hubble telescope. Comparing the two types of data will allow scientists to create even more precise algorithms for simulating galaxies inside supercomputers, and with further hardware improvements, those simulations will be able to be extended to galactic clusters, such as the "local cluster" to which the Milky Way belongs.