IBM and the US Department of Energy’s Oak Ridge National Laboratory (ORNL) announced the supercomputer Summit stating that it’s the fastest in the world. Peak performances are declared at 200 petaflops, which are eight times those of the previous ORNL supercomputer, called Titan, activated in October 2012 and for a few months the most powerful in its category. It’s a system designed for scientific applications with an eye to artificial intelligence.

Summit is the result of the efforts of IBM and ORNL about 4 years after the Department of Energy gave IBM the commission to build two new next-generation supercomputers. Being surpassed by the Chinese in the field of supercomputers has been a whammy for the Americans, but the race to increasingly powerful systems continues regardless of the rankings, for the need to use them for increasingly sophisticated applications.

To build Summit and its smaller “sister” Sierra at the Lawrence Livermore National Laboratory, IBM developed a new architecture that combines the high performance of POWER9 CPUs with NVIDIA GPUs, which are graphics card processors, optimized for use in artificial intelligence applications. These processors have been connected in order to ensure very high communication speeds but with components available to all companies because they’re the standard ones produced by IBM.

Summit employs 9.216 IBM 22-core POWER9 CPUs and 27.648 NVIDIA Tesla V100 GPUs. These processors are divided into 4,608 nodes where each one is made up of 2 POWER9 CPUs and 6 NVIDIA GPUs. The nodes are interconnected with a system that ensures a communication speed of 25 GBytes per second. To give an idea, the nodes could exchange several complete high-definition movies every second. As for memory, Summit has over 10 Petabytes of RAM and 250 Petabytes of mass storage. The entire supercomputer occupies a space equal to about two tennis courts. Its operating system is Red Hat Enterprise Linux.

Summit’s peak performance is declared at 200 petaflops, or 200 million billion calculations per second. For some scientific applications related to deep learning, a peak of 3.3 exaops could be achieved, meaning 3.3 billion billion mixed precision calculations per second. This is a technique that allows GPUs to achieve the same results more efficiently than the 64-bit ones performed by the Titan supercomputer. Summit has already reached 1.86 exaops for calculations in a research in the field of comparative genomics.

Scientific research of various kinds is the use that will be made of Summit. Problems will range from astrophysics simulations to materials research. There will be many biological researches, which will also include several health problems. For example, the US Department of Energy collaborates with the National Cancer Institute on cancer research.

At this point we just have to wait for the official tests of the supercomputer Summit to be published. If the statements about its performance will be confirmed, it will bring back to the US the title of the most powerful supercomputer after years in which Chinese supercomputers have been at the top of the Top 500 ranking. The new ranking will be released in the coming weeks and we can expect Summit to surpass Sunway TaihuLight.