Welcome to our NHR CREFLABS seminar today. Our guest today is Dr. Anton Kuzhevnikov.

He obtained his PhD from the Institute of Physics of Metals at the Ural Division of

the Russian Academy of Sciences in 2007. He later worked at the University of Tennessee in Knoxville

and then after that at ETH Zurich and this brought him to CSCS, the Swiss National

Supercomputing Center, where he's now working as a senior research software engineer.

He leads the development of what he's going to present about, namely the serious domain specific

for GPU, XRG, DFT ground states, for quantum expressor and similar in structure codes.

So we're looking forward to your talk, Anton. Take it away.

Okay, thank you, Georg, and thanks everyone for attendance. So I decided to a little bit not talk

only about serious, but also about the context where the development isn't by that. So I will

discuss briefly what is CSCS, what we do in terms of software development for scientific codes.

So then I will for sure talk about serious a little bit. I will present our status quo,

what we can do. And then also I will spend a little bit of time on presenting a few

lower level libraries that can also be of use to everyone who's doing referring structure

codes. So maybe you'll find that it's useful to you. And then we'll have questions and answers

session a little bit of time. So yeah, I work at CSCS. So I'm computational physicist by training.

So I do now a lot of other stuff at CSCS, but part of me is still with software development in

the electronic structure field. So we are located in two parts. So we have a big office is in Lugana.

I think this is where Georg visited us some time ago. And we also have from maybe five or six years,

we have a smaller office in Zurich in the northern part of Switzerland. And on the map on the left,

so you will see, so there's the small square is our office building. So with the shiny windows,

and in the back, so there's a great, great wall. And on top, you see this huge machine room.

This is where all our flagship machines are installed. And so we usually were like water

pipes come in, so they're cooling the machines and then they come out to Lugana.

And it's all like a lot of this a lot of hardware is installed there.

So we operate HPC flagship supercomputer for our research in Switzerland, but also for international

projects. And since quite some time, so we have a lot of yeah, so we deal with GP accelerated

machines. So we started in 2013. So when we brought to the floor our first big flagship

supercomputer based on K 20 X GPUs. So it was it's dying. And it was it was it bear this name for

until it's variant, which happened maybe like one month ago with the commission. So it's dying to

was hitting number six in top 100 in 2013 with 6.2 petaflop roughly. So then we made an upgrade.

So we upgraded it to P 100 based machine. So with 20 petaflops, and this was a big achievement,

because we hit numbers three and top 500. So we were like largest super largest HPC system in Europe.

And recently in 2024, we upgraded again our infrastructure. So we now have around 10,000

Grace Hopper 200 super chips assembled in several clusters. And so we are

about like 430 petaflops aggregated performance. And if you look at the specs of cars,

it's a tremendous increase from I know 1.7 petaflops for K 20 X to 34 petaflops of Grace

Hopper 267 with standard course if you like contain the tender force you have 67 petaflops

or cheap in double precision. In terms of memory bandwidth is also increased a lot. So it's an

amazing hardware. And you can imagine so for our users, it's it was a big, this is also a slide.

So from 2019, we are part of a Lumi community. So we have roughly 10% ownership in Lumi. And we have

we can send our users to Lumi and Lumi is basically so it has a CPU partition. But basically Lumi is

the huge GPU installation based on AMD K 20 K 250 X card. So what's important is that for our users,

it was the big and it's still a big challenge because they are usually the academia users,

they have their codes, they need to run, they don't care, not all of them care too much how

it can happen. So this, they need some help in porting and accelerating their codes to GPUs.

And this is like a long grow long walk. So we're working with our users. And for that, we have a

dedicated program established between CSS and I think, it's called platform for advanced

scientific computing. So it's, it's it funds around 14 projects on a three year cycles, and

it tries to bring both users from academia users who have their codes and ideas and then formulas,