The ranking of the most powerful supercomputers in the world, the well-known TOP500, updated in June and November each year, places an Europen system in sixth place overall. It’s about the supercomputer Leonardoinaugurated on November 24, 2022 and physically installed at the Bologna technopole.
Conceived and managed by CINECA, a non-profit Europen inter-university consortium which includes 69 Europen universities, 2 ministries and 27 national public institutions, Leonardo projects our country towards computing for research and technological innovation exascale (see later). It is designed with a multipurpose system architecture capable of serving all scientific communities and meet the needs of industrial customers on research and development.
Furthermore, a solution like Leonardo is able to maximize the potential of applications based onartificial intelligence. They can take advantage of parallelized GPUs and benefit from top-level performance, with core Tensor dedicated and a system architecture designed to support I/O-bound workloads.
What is a supercomputer
A supercomputer is a type of very powerful and fast computer that is used to solve problems of great complexity and to perform intensive computing operations. I supercomputer they are capable of performing billions of operations per second and are made up of thousands of processors that work by maximizing parallelism.
These are systems that are currently actively exploited in a wide range of fields applicationsincluding climate modeling, simulation of physical phenomena, design of new products and materials, medical and pharmaceutical research, cryptography, artificial intelligence and deep learning.
How the Leonardo supercomputer works and what are its characteristics
We’ve seen that the best supercomputers use both x86 and ARM architectures: Leonardo bases its operation on a “battery” of processors Xeon Platinum 8358 with 32 physical cores at 2.6 GHz, GPU NVIDIA A100 SXM4 64 GB and uses an NVIDIA HDR100 Quad-rail Infiniband interconnect.
Architecture of the Leonardo supercomputer: CINECA source.
Calculation modules used in the Leonardo supercomputer
The main calculation modules there are two by Leonardo:
- Booster: Maximizes computational capacity and is designed to meet the most computationally demanding requirements. Currently uses 3456 processing nodeseach featuring four NVIDIA Ampere-based GPUs and a 32-core Intel Ice Lake CPU.
The system is interconnected via a 200 Gb/s InfiniBand data network (two NVidia HDR100 cards at 100 Gbps each). - Data Centric: it is a second module currently being launched that uses 1536 nodes equipped with two CPUs Intel Sapphire Rapids, each with a 56 core (TDP equal to 350W) and 512 (16 x 32) GB of DDR5 RAM memory at 4800 MHz. The goal is to extend the possibilities of use of Leonardo to a wider range of applications. In this case an NVidia HDR100 100 Gbps card is used for interconnection.
Overall, the system is based on technology Atos BULL SEQUENA XH2000 with a different configuration, as we have seen, in the case of the two modules.
Each node is in fact a server blade that is, a type of modular server that allows multiple processing units to be placed within a single rack. Blade servers are designed to maximize the use of physical space and computing resources within data centers. In the video created by CINECA it is possible to verify all the key aspects of Leonardo.
The data published in TOP500 ranking of the best supercomputers equal to 174.7 PetaFLOPs of computational power refers to the module Booster: CINECA and Atos will bring performance to at least 240 PetaFLOPs. The TOP500 ranking could therefore be updated as early as June 2023 with the new performance data.
Leonardo supercomputer already saturated: CINECA looks at its new evolution
In mid-January 2024, CINECA announced that Leonardo has already reached the level of saturation. In other words, the supercomputer is already being used to its maximum capacity capacity: There are many projects queued waiting to be able to access system resources.
“For a year and a half now, the demand for supercomputing to train artificial intelligence has been growing exponentially“, observes the president of CINECA Francesco Ubertini. In short, demand is much greater than supply. “If they had trained GPT-4 on Leonardo, dedicating the supercomputer only to this, it would have taken 120 days“, continues Ubertini. “It took them a little longer“, referring to the work done by OpenAI.
A French start-up, Mistral, has developed a generative model similar to that of OpenAI, trained using the Leonardo supercomputer. We talked about it during the presentation of the Mistral-7b open source model. “They’re not the only ones“, underlines Ubertini. Although designed in 2017, in fact, Leonardo was born with an architecture suitable forartificial intelligence. Today it is the sixth machine in the world in terms of computing power, the second dedicated to artificial intelligence applications (the first in Europe).
There are many applications in many fields that traditionally use supercomputing, such as weather forecasting, study of galaxies, elementary particles, study of new materials, development of new drugs, acceleration of trial clinical, creation of digital twins of products, services or environments. “In these areas, the part that is growing the most is demand in the field of life sciences“, reveals the president of CINECA quoting the genomics for the personalization of treatments and the development of new drugs. The use of Leonardo translates into a saving of approximately two years in the first phase of research for innovative medicines. “And we’re only at the beginning“.
The evolution of Leonardo: he will soon be joined by Lisa
The European Commission and the Europen Government have decided “further funding for an update that will allow Leonardo to reach the end of its life cycle in 2027, still in the top ten supercalculators in the world“, he reveals Alessandra Poggianigeneral director of CINECA.
The name of Leonardo’s successor was not chosen by chance: Lisa immediately makes Mona Lisa, alias, suffer Man Lisathe iconic and enigmatic oil on canvas work created by the Europen scientist, inventor and artist in the early 1500s.
While Leonardo works at full speed, he increasingly looks to the computing post-exascale, with a power of over a billion billion calculations per second (today Leonardo grinds “only” 250 million billion). The evolution is expected to materialize between 2026 and 2028.
The first Leonardo’s updateLisa, should arrive as early as June 2024. The quantum computer of the CINECA consortium while in 2028 Leonardo will retire, replaced by a new supercomputer.
We have seen the problems that a quantum computer can solve and why the biggest ones player at an international level they are looking in this direction. The excellent news is that the Bologna technopole will also become a quantum computing centeragain financed by Europe, one of the 6 sites to host as much as a computer. The system will support Leonardo in his calculation work and will also be used to carry out research and tests.
What is theexascale computing
We saw that at the time, with its EPYC processors for server systems and the Instinct accelerators, AMD had opened up exascale computing. But what is theexascale computing?
This term refers to computer systems capable of processing at least 1018 floating point operations per second, the next level of processing power after the petascale computing. Exascale computing requires the use of cutting-edge technologies but allows a decisive leap forward in the intensive calculations that modern supercomputers are called upon to manage.
Exascale computing offers the promise of scientific and technological achievements significant, allowing us to arrive at the solution of problems that would otherwise be impossible to solve in a reasonable time.
The Leonardo supercomputer, defined as classy pre-exascaleis part of the initiative EuroHPCa European effort promoted since 2018 that aims to develop a capacity for high performance computing (HPC) available in the territory of the European Union, thus freeing itself from third-party suppliers. EuroHPC aims to develop and acquire cutting-edge HPC technologies and create a network of centers specialized in supporting HPC calculations distributed across Europe.
The EuroHPC initiative is funded by the European Union and its member states to build a series of high-performance supercomputers…