Page 51 - EE Times Europe Magazine – June 2024

EE|Times EUROPE 51

BSC’s Girona on MareNostrum’s Evolution and Delivering World-Class Compute from Barcelona

MareNostrum 5 supercomputer (Source: BSC-CNS)

the major objective is to support science and For example, it will run simulations for new The second challenge is that the systems
innovation in Spain and the rest of Europe. aircraft designs. are more complex today, requiring more
Many of the users do applied research in We’re seeing increasing demand in several specialized components. Supercomputers
collaboration with industrial players to other domains. This includes energy in are water-cooled, which means they need
develop new methods and algorithms to make general and nuclear fusion in particular. hoses and pipes that withstand high tem-
advances in science and technology. For example, for wind farms, you have to peratures and need to be certified. And the
We’ve supported many applications that model the behavior of the wind, which is very components are very dense, which makes it
achieve breakthroughs in science and tech- complex; and for nuclear fusion, you have difficult to build the bigger system and to
nology. For example, we dedicated parts of to analyze the behavior of plasma inside remove and replace parts.
the system to help develop vaccines for the the reactor. The demand also includes large Today’s supercomputers are unique. The
coronavirus and to solve distribution prob- language models and physical and chemical cost of the systems is very high, and the
lems to get the vaccines to people. Another simulations and modelization at the highest supplies are very scarce. MareNostrum 1
example is the work the ESA [European Space level. For example, personalized medicine was based on the PowerPC 970, which was
Agency] has been doing to build a 3D map of requires more complex analysis to adapt the processor used for Apple Macintosh
the universe. treatment to a specific patient. To really ana- at the time. It was very easy to buy those
By increasing the computing power, system lyze the behavior of an individual patient’s general-purpose processors.
memory and number of cores, heart or to design drugs that are tailored for Today, we’re using Intel Sapphire Rapids
MareNostrum 5 will help solve more and specific patients, exascale systems will be with 56 cores, which, because it isn’t sold
more complex problems. For example, climate required. to the general market, is in limited supply—
change simulations will be able to have and the same is true for the Hopper 100.
higher resolution, moving from representing EE TIMES EUROPE: What are some of the We have a total of 4,480 H100s, which is a
phenomena on spatial scales of hundreds of obstacles slowing down the evolution of big part of what’s available on the market.
kilometers to include processes occurring on supercomputers? When one is damaged and you need a new
scales of a few kilometers, making predictions Girona: Today, it’s really hard to get a one, it might not be easy to get a replace-
much more accurate and reliable. supercomputer up to the state of production ment quickly.
It will also be possible to tackle much more on time. From my experience with systems in Cables are also a challenge. The number
complex problems of artificial intelligence the past, we set a date and managed to have of cables and the different lengths make it
and big data analysis. For example, it will be the system in operation on that date. That is very challenging to keep what you need in
possible to generate massive language models nearly impossible now. stock. You have to find a reliable supplier to
by training much larger neural networks with The first challenge is that materials and get the parts you need.
hundreds of billions of parameters, using capacities are lacking. Supply chains still Another obstacle that slows things down is
infinitely larger datasets than today. haven’t recovered since the pandemic, and the software that’s used on supercomputers.
MareNostrum 5 specifically targets different events around the world continue We have been too focused on maintaining
European medical research, which includes to produce cascading effects. We saw this legacy code, when instead we should be
the design of new drugs, the development of recently with the Suez Canal crisis and with rethinking the ways we solve problems. If
vaccines and the simulation of virus propa- the earthquake in Taiwan in early April, when you could change some of the algorithms to
gation. We also expect it to be an essential TSMC and other manufacturers temporarily take advantage of the newer architectures,
tool for materials science and engineering. stopped production of semiconductors. you could create new opportunities. ■

www.eetimes.eu | JUNE 2024